Digital Control
Engineering
Analysis and Design
Second Edition
M. Sami Fadali
Antonio Visioli
AMSTERDAM • BOSTON • HEIDELBERG • LONDON
NEW YORK • OXFORD • PARIS • SAN DIEGO
SAN FRANCISCO • SINGAPORE • SYDNEY • TOKYO
Academic Press is an imprint of Elsevier
Academic Press is an imprint of Elsevier
225 Wyman Street, Waltham, MA 02451, USA
The Boulevard, Langford Lane, Kidlington, Oxford, OX5 1GB, UK
Copyright r 2013 Elsevier Inc. All rights reserved
No part of this publication may be reproduced or transmitted in any form or by any means, electronic
or mechanical, including photocopying, recording, or any information storage and retrieval system,
without permission in writing from the publisher. Details on how to seek permission, further
information about the Publisher’s permissions policies and our arrangements with organizations
such as the Copyright Clearance Center and the Copyright Licensing Agency, can be found at our
website: www.elsevier.com/permissions.
This book and the individual contributions contained in it are protected under copyright by the
Publisher (other than as may be noted herein).
Notices
Knowledge and best practice in this field are constantly changing. As new research and
experience broaden our understanding, changes in research methods, professional practices, or
medical treatment may become necessary.
Practitioners and researchers must always rely on their own experience and knowledge in
evaluating and using any information, methods, compounds, or experiments described herein.
In using such information or methods they should be mindful of their own safety and the safety
of others, including parties for whom they have a professional responsibility.
To the fullest extent of the law, neither the Publisher nor the authors, contributors, or editors,
assume any liability for any injury and/or damage to persons or property as a matter of products
liability, negligence or otherwise, or from any use or operation of any methods, products, instructions,
or ideas contained in the material herein.
Library of Congress Cataloging-in-Publication Data
Fadali, M. Sami.
Digital control engineering : analysis and design / M. Sami Fadali, Antonio Visioli. Second edition.
pages cm
Includes bibliographical references and index.
ISBN 978-0-12-394391-0 (hardback)
1. Digital control systems. I. Visioli, Antonio. II. Title.
TJ223.M53F33 2013
629.80 9dc23
2012021488
British Library Cataloguing-in-Publication Data
A catalogue record for this book is available from the British Library
For information on all Academic Press publications
visit our website at http://store.elsevier.com
Printed in the United States of America
12 13 14 9 8 7 6 5 4 3 2 1
Preface
Approach
Control systems are an integral part of everyday life in today’s society. They
control our appliances, our entertainment centers, our cars, and our office environments; they control our industrial processes and our transportation systems; they
control our exploration of land, sea, air, and space. Almost all of these applications
use digital controllers implemented with computers, microprocessors, or digital
electronics. Every electrical, chemical, or mechanical engineering senior or graduate student should therefore be familiar with the basic theory of digital
controllers.
This text is designed for a senior or combined senior/graduate-level course in
digital controls in departments of mechanical, electrical, or chemical engineering.
Although other texts are available on digital controls, most do not provide a satisfactory format for a senior/graduate-level class. Some texts have very few examples to support the theory, and some were written before the wide availability of
computer-aided-design (CAD) packages. Others use CAD packages in certain
ways but do not fully exploit their capabilities. Most available texts are based on
the assumption that students must complete several courses in systems and control
theory before they can be exposed to digital control. We disagree with this
assumption, and we firmly believe that students can learn digital control after a
one-semester course covering the basics of analog control. As with other topics
that started at the graduate level—linear algebra and Fourier analysis to name a
few—the time has come for digital control to become an integral part of the
undergraduate curriculum.
Features
To meet the needs of the typical senior/graduate-level course, this text includes
the following features.
Numerous examples
The book includes a large number of examples. Typically, only one or two examples can be covered in the classroom because of time limitations. The student can
use the remaining examples for self-study. The experience of the authors is that
students need more examples to experiment with so as to gain a better understanding of the theory. The examples are varied to bring out subtleties of the
theory that students may overlook.
xi
xii
Preface
Extensive use of CAD packages
The book makes extensive use of CAD packages. It goes beyond the occasional
reference to specific commands to the integration of these commands into the
modeling, design, and analysis of digital control systems. For example, root locus
design procedures given in most digital control texts are not CAD procedures and
instead emphasize paper-and-pencil design. The use of CAD packages, such as
MATLABs, frees students from the drudgery of mundane calculations and
allows them to ponder more subtle aspects of control system analysis and design.
The availability of a simulation tool like Simulinks allows the student to simulate closed-loop control systems, including aspects neglected in design such as
nonlinearities and disturbances.
Coverage of background material
The book itself contains review material from linear systems and classical control.
Some background material is included in the appendices that could either be
reviewed in class or consulted by the student as necessary. The review material,
which is often neglected in digital control texts, is essential for the understanding
of digital control system analysis and design. For example, the behavior of discrete-time systems in the time domain and in the frequency domain is a standard
topic in linear systems texts but often receives brief coverage. Root locus design
is almost identical for analog systems in the s-domain and digital systems in the
z-domain. The topic is covered much more extensively in classical control texts
and inadequately in digital control texts. The digital control student is expected to
recall this material or rely on other sources. Often, instructors are obliged to compile their own review materials, and the continuity of the course is adversely
affected.
Inclusion of advanced topics
In addition to the basic topics required for a one-semester senior/graduate class,
the text includes some advanced material to make it suitable for an introductory
graduate-level class or for two quarters at the senior/graduate level. We would
also hope that the students in a single-semester course would acquire enough
background and interest to read the additional chapters on their own. Examples of
optional topics are statespace methods, which may receive brief coverage in a
one-semester course, and nonlinear discrete-time systems, which may not be
covered.
Standard mathematics prerequisites
The mathematics background required for understanding most of the book
does not exceed what can be reasonably expected from the average electrical,
chemical, or mechanical engineering senior. This background includes three
Preface
semesters of calculus, differential equations, and basic linear algebra. Some
texts on digital control require more mathematical maturity and are therefore
beyond the reach of the typical senior. On the other hand, the text does include
optional topics for the more advanced student. The rest of the text does not
require knowledge of this optional material so that it can be easily skipped if
necessary.
Senior system theory prerequisites
The control and system theory background required for understanding the book
does not exceed material typically covered in one semester of linear systems and
one semester of control systems. Thus, students should be familiar with Laplace
transforms, the frequency domain, and the root locus. They need not be familiar
with the behavior of discrete-time systems in the frequency and time domain or
have extensive experience with compensator design in the s-domain. For an audience with an extensive background in these topics, some topics can be skipped
and the material can be covered at a faster rate.
Coverage of theory and applications
The book has two authors: the first is primarily interested in control theory and
the second is primarily interested in practical applications and hardware implementation. Even though some control theorists have sufficient familiarity with
practical issues such as hardware implementation and industrial applications to
touch on the subject in their texts, the material included is often deficient because
of the rapid advances in the area and the limited knowledge that theorists have of
the subject.
New to this edition
We made several important changes and added material to the second edition:
1. We added a brief introduction to Simulink simulation of discrete-time systems
to Chapter 3.
2. We moved the explanation of the bilinear transform to Chapter 4, where the
bilinear transform is first introduced, from Chapter 6.
3. We added closed-loop Ziegler-Nichols design to Chapter 5.
4. We added pole-zero matching to Chapter 6. This is a simple design approach
that was used in some examples but was not included in the first edition.
5. We have improved the explanation of the direct control design (Section 6.6)
and of the finite settling time design (Section 6.7).
6. We added the Hankel realization to Chapter 8 to provide a systematic method
for multi-input-multi-output system realization. Because this material is based
on the singular value decomposition, a section on the singular value
decomposition was added to Appendix III.
xiii
xiv
Preface
7. In the first edition, the Hamiltonian system was included, but the significance
of its eigenstructure was not discussed. We added a section on the
eigenstructure of the Hamiltonian system to Chapter 10.
8. The first edition did not include a discussion of the stability of the response of
the system to an external input. We added input-output stability and the circle
criterion to Chapter 11.
9. We added 23 new problems, including several new computer exercises.
It became clear to the first author that to have a suitable text for his course
and similar courses, he needed to find a partner to satisfactorily complete the text.
He gradually collected material for the text and started looking for a qualified and
interested partner. Finally, he found a co-author who shared his interest in digital
control and the belief that it can be presented at a level amenable to the average
undergraduate engineering student.
For many years, Dr. Antonio Visioli has been teaching an introductory and a
laboratory course on automatic control, as well as a course on control systems
technology. Further, his research interests are in the fields of industrial regulators
and robotics. Although he contributed to the material presented throughout the
text, his major contribution was adding material related to the practical design
and implementation of digital control systems. This material is rarely covered in
control systems texts but is an essential prerequisite for applying digital control
theory in practice.
The text is written to be as self-contained as possible. However, the reader is
expected to have completed a semester of linear systems and classical control.
Throughout the text, extensive use is made of the numerical computation and
computer-aided-design package MATLAB. As with all computational tools, the
enormous capabilities of MATLAB are no substitute for a sound understanding
of the theory presented in the text. As an example of the inappropriate use of supporting technology, we recall the story of the driver who followed the instructions
of his GPS system and drove into the path of an oncoming train!1 The reader
must use MATLAB as a tool to support the theory without blindly accepting its
computational results.
Organization of text
The text begins with an introduction to digital control and the reasons for its popularity. It also provides a few examples of applications of digital control from the
engineering literature.
1
The story was reported in the Chicago Sun-Times, on January 4, 2008. The driver, a computer
consultant, escaped just in time before the train slammed into his car at 60 mph in Bedford Hills,
New York.
Preface
Chapter 2 considers discrete-time models and their analysis using the
z-transform. We review the z-transform, its properties, and its use to solve difference equations. The chapter also reviews the properties of the frequency
response of discrete-time systems. After a brief discussion of the sampling theorem, we are able to provide rules of thumb for selecting the sampling rate for a
given signal or for given system dynamics. This material is often covered in linear systems courses, and much of it can be skipped or covered quickly in a digital
control course. However, the material is included because it serves as a foundation for much of the material in the text.
Chapter 3 derives simple mathematical models for linear discrete-time systems. We derive models for the analog-to-digital converter (ADC), the digital-toanalog converter (DAC), and an analog system with a DAC and an ADC. We
include systems with time delays that are not an integer multiple of the sampling
period. These transfer functions are particularly important because many applications include an analog plant with DAC and ADC. Nevertheless, there are situations where different configurations are used. We therefore include an analysis of
a variety of configurations with samplers. We also characterize the steady-state
tracking error of discrete-time systems and define error constants for the unity
feedback case. These error constants play an analogous role to the error constants
for analog systems. Using our analysis of more complex configurations, we are
able to obtain the error due to a disturbance input.
In Chapter 4, we present stability tests for input-output systems. We examine
the definitions of input-output stability and internal stability and derive conditions for each. By transforming the characteristic polynomial of a discrete-time
system, we are able to test it using the standard Routh-Hurwitz criterion for
analog systems. We use the Jury criterion, which allows us to directly test the
stability of a discrete-time system. Finally, we present the Nyquist criterion for
the z-domain and use it to determine closed-loop stability of discrete-time
systems.
Chapter 5 introduces analog s-domain design of proportional (P), proportionalplus-integral (PI), proportional-plus-derivative (PD), and proportional-plusintegral-plus-derivative (PID) control using MATLAB. We use MATLAB as an
integral part of the design process, although many steps of the design can be competed using a scientific calculator. It would seem that a chapter on analog design
does not belong in a text on digital control. This is false. Analog control can be
used as a first step toward obtaining a digital control. In addition, direct digital
control design in the z-domain is similar in many ways to s-domain design.
Digital controller design is topic of Chapter 6. It begins with proportional control design then examines digital controllers based on analog design. The direct
design of digital controllers is considered next. We consider root locus design in
the z-plane for PI and PID controllers. We also consider a synthesis approach due
to Ragazzini that allows us to specify the desired closed-loop transfer function.
As a special case, we consider the design of deadbeat controllers that allow us to
xv
xvi
Preface
exactly track an input at the sampling points after a few sampling points. For
completeness, we also examine frequency response design in the w-plane. This
approach requires more experience because values of the stability margins must
be significantly larger than in the more familiar analog design. As with analog
design, MATLAB is an integral part of the design process for all digital control
approaches.
Chapter 7 covers statespace models and statespace realizations. First, we
discuss analog statespace equations and their solutions. We include nonlinear
analog equations and their linearization to obtain linear statespace equations.
We then show that the solution of the analog state equations over a sampling
period yields a discrete-time statespace model. Properties of the solution of the
analog state equation can thus be used to analyze the discrete-time state equation.
The discrete-time state equation is a recursion for which we obtain a solution by
induction. In Chapter 8, we consider important properties of statespace models:
stability, controllability, and observability. As in Chapter 4, we consider internal
stability and input-output stability, but the treatment is based on the properties of
the statespace model rather than those of the transfer function. Controllability is
a property that characterizes our ability to drive the system from an arbitrary
initial state to an arbitrary final state in finite time. Observability characterizes
our ability to calculate the initial state of the system using its input and output
measurements. Both are structural properties of the system that are independent
of its stability. Next, we consider realizations of discrete-time systems. These are
ways of implementing discrete-time systems through their statespace equations
using summers and delays.
Chapter 9 covers the design of controllers for statespace models. We show
that the system dynamics can be arbitrarily chosen using state feedback if the
system is controllable. If the state is not available for feedback, we can design a
state estimator or observer to estimate it from the output measurements. These
are dynamic systems that mimic the system but include corrective feedback to
account for errors that are inevitable in any implementation. We give two types
of observers. The first is a simpler but more computationally costly full-order
observer that estimates the entire state vector. The second is a reduced-order
observer with the order reduced by virtue of the fact that the measurements are
available and need not be estimated. Either observer can be used to provide an
estimate of the state for feedback control, or for other purposes. Control schemes
based on state estimates are said to use observer state feedback.
Chapter 10 deals with the optimal control of digital control systems. We
consider the problem of unconstrained optimization, followed by constrained optimization, then generalize to dynamic optimization as constrained by the system
dynamics. We are particularly interested in the linear quadratic regulator where
optimization results are easy to interpret and the prerequisite mathematics background is minimal. We consider both the finite time and steady-state regulator
and discuss conditions for the existence of the steady-state solution. The first
10 chapters are mostly restricted to linear discrete-time systems. Chapter 11
Preface
examines the far more complex behavior of nonlinear discrete-time systems. It
begins with equilibrium points and their stability. It shows how equivalent discrete-time models can be easily obtained for some forms of nonlinear analog systems using global or extended linearization. It provides stability theorems and
instability theorems using Lyapunov stability theory. The theory gives sufficient
conditions for nonlinear systems, and failure of either the stability or instability
tests is inconclusive. For linear systems, Lyapunov stability yields necessary and
sufficient conditions. Lyapunov stability theory also allows us to design controllers by selecting a control that yields a closed-loop system that meets the
Lyapunov stability conditions. For the classes of nonlinear systems for which
extended linearization is straightforward, linear design methodologies can yield
nonlinear controllers.
Chapter 12 deals with practical issues that must be addressed for the successful implementation of digital controllers. In particular, the hardware and software
requirements for the correct implementation of a digital control system are
analyzed. We discuss the choice of the sampling frequency in the presence of
antialiasing filters and the effects of quantization, rounding, and truncation errors.
We also discuss bumpless switching from automatic to manual control, avoiding
discontinuities in the control input. Our discussion naturally leads to approaches for
the effective implementation of a PID controller. Finally, we consider nonuniform
sampling, where the sampling frequency is changed during control operation, and
multirate sampling, where samples of the process outputs are available at a slower
rate than the controller sampling rate.
Supporting material
The following resources are available to instructors adopting this text for use in
their courses. Please visit textbooks.elsevier.com to register for access to these
materials:
Instructor Solutions Manual. Fully typeset solutions to the end-of-chapter
problems in the text.
PowerPoints Images. Electronic images of the figures and tables from the
book, useful for creating lectures.
xvii
Contents
Preface ......................................................................................................................xi
CHAPTER 1
Introduction to Digital Control .............................................1
1.1 Why digital control?......................................................................2
1.2 The structure of a digital control system......................................2
1.3 Examples of digital control systems .............................................3
1.3.1 Closed-loop drug delivery system .......................................3
1.3.2 Computer control of an aircraft turbojet engine .................4
1.3.3 Control of a robotic manipulator .........................................4
Resources...............................................................................................6
Problems ................................................................................................7
CHAPTER 2
Discrete-Time Systems ..........................................................9
2.1 Analog systems with piecewise constant inputs...........................9
2.2 Difference equations ...................................................................11
2.3 The z-transform ...........................................................................12
2.3.1 z-Transforms of standard discrete-time signals.................13
2.3.2 Properties of the z-transform .............................................15
2.3.3 Inversion of the z-transform...............................................19
2.3.4 The final value theorem .....................................................28
2.4 Computer-aided design ...............................................................29
2.5 z-Transform solution of difference equations.............................31
2.6 The time response of a discrete-time system .............................32
2.6.1 Convolution summation.....................................................32
2.6.2 The convolution theorem ...................................................34
2.7 The modified z-transform............................................................37
2.8 Frequency response of discrete-time systems.............................39
2.8.1 Properties of the frequency response of
discrete-time systems .........................................................42
2.8.2 MATLAB commands for the discrete-time
frequency response.............................................................44
2.9 The sampling theorem.................................................................45
2.9.1 Selection of the sampling frequency .................................46
Resources.............................................................................................49
Problems ..............................................................................................49
Computer exercises .............................................................................52
iii
iv
Contents
CHAPTER 3
Modeling of Digital Control Systems ..............................55
3.1
3.2
3.3
3.4
3.5
ADC model .................................................................................55
DAC model .................................................................................56
The transfer function of the ZOH ..............................................57
Effect of the sampler on the transfer function of a cascade......58
DAC, analog subsystem, and ADC combination
transfer function ..........................................................................61
3.6 Systems with transport lag..........................................................69
3.7 The closed-loop transfer function...............................................71
3.8 Analog disturbances in a digital system.....................................74
3.9 Steady-state error and error constants ........................................75
3.9.1 Sampled step input ............................................................77
3.9.2 Sampled ramp input ..........................................................77
3.10 MATLAB commands .................................................................79
3.10.1 MATLAB ........................................................................79
3.10.2 Simulink ..........................................................................80
Resources.............................................................................................85
Problems ..............................................................................................85
Computer exercises .............................................................................89
CHAPTER 4
Stability of Digital Control Systems ................................91
4.1 Definitions of stability ................................................................91
4.2 Stable z-domain pole locations...................................................93
4.3 Stability conditions .....................................................................94
4.3.1 Asymptotic stability ..........................................................94
4.3.2 BIBO stability ...................................................................95
4.3.3 Internal stability.................................................................98
4.4 Stability determination..............................................................101
4.4.1 MATLAB ........................................................................101
4.4.2 Routh-Hurwitz criterion ..................................................102
4.5 Jury test .....................................................................................104
4.6 Nyquist criterion .......................................................................109
4.6.1 Phase margin and gain margin........................................114
Resources...........................................................................................123
Problems ............................................................................................123
Computer exercises ...........................................................................125
CHAPTER 5
Analog Control System Design........................................127
5.1 Root locus .................................................................................127
5.2 Root locus using MATLAB .....................................................132
5.3 Design specifications and the effect of gain variation.............132
Contents
5.4 Root locus design ......................................................................135
5.4.1 Proportional control .........................................................137
5.4.2 PD control ........................................................................138
5.4.3 PI control..........................................................................147
5.4.4 PID control .......................................................................153
5.5 Empirical tuning of PID controllers .........................................156
Resources...........................................................................................161
Problems ............................................................................................161
Computer exercises ...........................................................................163
CHAPTER 6
Digital Control System Design .........................................165
6.1 z-Domain root locus ..................................................................165
6.2 z-Domain digital control system design ...................................168
6.2.1 z-Domain contours ...........................................................171
6.2.2 Proportional control design in the z-domain ...................175
6.3 Digital implementation of analog controller design.................180
6.3.1 Differencing methods.......................................................181
6.3.2 Pole-zero matching ..........................................................183
6.3.3 Bilinear transformation ....................................................186
6.3.4 Empirical digital PID controller tuning...........................199
6.4 Direct z-domain digital controller design .................................200
6.5 Frequency response design .......................................................205
6.6 Direct control design .................................................................213
6.7 Finite settling time design.........................................................218
Resources...........................................................................................230
Problems ............................................................................................230
Computer exercises ...........................................................................233
CHAPTER 7
StateSpace Representation ..........................................235
7.1 State variables ...........................................................................235
7.2 Statespace representation .......................................................238
7.2.1 Statespace representation in MATLAB........................240
7.2.2 Linear versus nonlinear statespace equations...............240
7.3 Linearization of nonlinear state equations................................243
7.4 The solution of linear statespace equations...........................246
7.4.1 The Leverrier algorithm...................................................251
7.4.2 Sylvester’s expansion.......................................................255
7.4.3 The state-transition matrix for a diagonal state matrix...257
7.4.4 Real form for complex conjugate eigenvalues................262
7.5 The transfer function matrix .....................................................264
7.5.1 MATLAB commands.......................................................265
v
vi
Contents
7.6 Discrete-time statespace equations ........................................266
7.6.1 MATLAB commands for discrete-time
statespace equations ......................................................269
7.6.2 Complex conjugate eigenvalues .................................... 269
7.7 Solution of discrete-time statespace equations......................271
7.7.1 z-Transform solution of discrete-time state equations ...... 272
7.8 z-Transfer function from statespace equations ......................277
7.8.1 z-Transfer function in MATLAB.....................................279
7.9 Similarity Transformation .........................................................279
7.9.1 Invariance of transfer functions and
characteristic equations ....................................................282
Resources...........................................................................................283
Problems ............................................................................................283
Computer exercises ...........................................................................289
CHAPTER 8
Properties of StateSpace Models ...............................293
8.1 Stability of statespace realizations.........................................294
8.1.1 Asymptotic stability.......................................................294
8.1.2 BIBO stability ..................................................................297
8.2 Controllability and stabilizability............................................301
8.2.1 MATLAB commands for controllability testing.............307
8.2.2 Controllability of systems in normal form ......................308
8.2.3 Stabilizability.................................................................309
8.3 Observability and detectability .................................................313
8.3.1 MATLAB commands.......................................................316
8.3.2 Observability of systems in normal form........................317
8.3.3 Detectability...................................................................317
8.4 Poles and zeros of multivariable systems.................................319
8.4.1 Poles and zeros from the transfer function matrix..........320
8.4.2 Zeros from statespace models ....................................323
8.5 Statespace realizations ...........................................................325
8.5.1 Controllable canonical realization ...................................326
8.5.2 Controllable form in MATLAB .....................................330
8.5.3 Parallel realization .........................................................331
8.5.4 Observable form...............................................................336
8.6 Duality........................................................................................... 338
8.7 Hankel realization ..................................................................339
Resources...........................................................................................343
Problems ............................................................................................344
Computer exercises ...........................................................................349
Contents
CHAPTER 9
State Feedback Control .....................................................351
9.1 State and output feedback .........................................................351
9.2 Pole placement...........................................................................353
9.2.1 Pole placement by transformation to
controllable form..............................................................356
9.2.2 Pole placement using a matrix polynomial .....................357
9.2.3 Choice of the closed-loop eigenvalues............................359
9.2.4 MATLAB commands for pole placement.......................364
9.2.5 Pole placement for multi-input systems ..........................364
9.2.6 Pole placement by output feedback.................................367
9.3 Servo problem............................................................................367
9.4 Invariance of system zeros ........................................................372
9.5 State estimation..........................................................................374
9.5.1 Full-order observer...........................................................374
9.5.2 Reduced-order observer ...................................................377
9.6 Observer state feedback.............................................................380
9.6.1 Choice of observer eigenvalues .......................................383
9.7 Pole assignment using transfer functions..................................389
Resources...........................................................................................393
Problems ............................................................................................393
Computer exercises ...........................................................................397
CHAPTER 10 Optimal Control .....................................................................399
10.1 Optimization ..............................................................................399
10.1.1 Unconstrained optimization...........................................400
10.1.2 Constrained optimization...............................................402
10.2 Optimal control..........................................................................404
10.3 The linear quadratic regulator ...................................................409
10.3.1 Free final state ...............................................................410
10.4 Steady-state quadratic regulator ................................................419
10.4.1 Output quadratic regulator .............................................420
10.4.2 MATLAB solution of the steady-state
regulator problem .........................................................421
10.4.3 Linear quadratic tracking controller ..............................423
10.5 Hamiltonian system ...................................................................426
10.5.1 Eigenstructure of the Hamiltonian matrix.....................429
Resources .............................................................................................433
Problems ..............................................................................................433
Computer exercises..............................................................................436
vii
viii
Contents
CHAPTER 11 Elements of Nonlinear Digital Control Systems ........439
11.1 Discretization of nonlinear systems ..........................................439
11.1.1 Extended linearization by input redefinition ................440
11.1.2 Extended linearization by input and
state redefinition ............................................................442
11.1.3 Extended linearization by output differentiation ..........443
11.1.4 Extended linearization using matching conditions .......445
11.2 Nonlinear difference equations .................................................447
11.2.1 Logarithmic transformation...........................................448
11.3 Equilibrium of nonlinear discrete-time systems .......................448
11.4 Lyapunov stability theory..........................................................450
11.4.1 Lyapunov functions .......................................................450
11.4.2 Stability theorems ..........................................................452
11.4.3 Rate of convergence ......................................................454
11.4.4 Lyapunov stability of linear systems ............................454
11.4.5 MATLAB.......................................................................457
11.4.6 Lyapunov’s linearization method..................................458
11.4.7 Instability theorems .......................................................459
11.4.8 Estimation of the domain of attraction .........................461
11.5 Stability of analog systems with digital control .......................463
11.6 State plane analysis ...................................................................465
11.7 Discrete-time nonlinear controller design.................................470
11.7.1 Controller design using extended linearization.............470
11.7.2 Controller design based on Lyapunov
stability theory ...............................................................473
11.8 Input-output stability and the small gain theorem....................474
11.8.1 Absolute stability ...........................................................481
Resources .............................................................................................485
Problems ..............................................................................................485
Computer exercises..............................................................................489
CHAPTER 12 Practical Issues ...................................................................491
12.1 Design of the hardware and software architecture ...................491
12.1.1 Software requirements ...................................................491
12.1.2 Selection of ADC and DAC..........................................494
12.2 Choice of the sampling period ..................................................495
12.2.1 Antialiasing filters .........................................................495
12.2.2 Effects of quantization errors ........................................498
12.2.3 Phase delay introduced by the ZOH .............................503
Contents
12.3 Controller structure....................................................................504
12.4 PID control.................................................................................507
12.4.1 Filtering the derivative action........................................507
12.4.2 Integrator windup ...........................................................509
12.4.3 Bumpless transfer between manual
and automatic mode .......................................................512
12.4.4 Incremental form............................................................515
12.5 Sampling period switching ........................................................516
12.5.1 MATLAB commands.....................................................519
12.5.2 Dual-rate control ............................................................526
Resources .............................................................................................528
Problems ..............................................................................................529
Computer exercises..............................................................................530
APPENDIX I
Table of Laplace and z-transforms ...............................533
APPENDIX II
Properties of the z-transform ..........................................535
APPENDIX III Review of Linear Algebra .................................................537
Index ......................................................................................................................565
ix
CHAPTER
Introduction to Digital Control
1
OBJECTIVES
After completing this chapter, the reader will be able to do the following:
1. Explain the reasons for the popularity of digital control systems.
2. Draw a block diagram for digital control of a given analog control system.
3. Explain the structure and components of a typical digital control system.
In most modern engineering systems, it is necessary to control the evolution
with time of one or more of the system variables. Controllers are required to
ensure satisfactory transient and steady-state behavior for these engineering systems. To guarantee satisfactory performance in the presence of disturbances and
model uncertainty, most controllers in use today employ some form of negative
feedback. A sensor is needed to measure the controlled variable and compare its
behavior to a reference signal. Control action is based on an error signal defined
as the difference between the reference and the actual values.
The controller that manipulates the error signal to determine the desired
control action has classically been an analog system, which includes electrical,
fluid, pneumatic, or mechanical components. These systems all have analog
inputs and outputs (i.e., their input and output signals are defined over a continuous time interval and have values that are defined over a continuous range
of amplitudes). In the past few decades, analog controllers have often been
replaced by digital controllers whose inputs and outputs are defined at discrete
time instances. The digital controllers are in the form of digital circuits, digital
computers, or microprocessors.
Intuitively, one would think that controllers that continuously monitor the
output of a system would be superior to those that base their control on sampled
values of the output. It would seem that control variables (controller outputs) that
change continuously would achieve better control than those that change periodically. This is in fact true! Had all other factors been identical for digital and
analog control, analog control would be superior to digital control. What, then,
is the reason behind the change from analog to digital that has occurred over the
past few decades?
1
2
CHAPTER 1 Introduction to Digital Control
1.1 Why digital control?
Digital control offers distinct advantages over analog control that explain its popularity. Here are some of its many advantages:
Accuracy. Digital signals are represented in terms of zeros and ones with
typically 12 bits or more to represent a single number. This involves a very
small error as compared to analog signals, where noise and power supply drift
are always present.
Implementation errors. Digital processing of control signals involves
addition and multiplication by stored numerical values. The errors that result
from digital representation and arithmetic are negligible. By contrast, the
processing of analog signals is performed using components such as resistors
and capacitors with actual values that vary significantly from the nominal
design values.
Flexibility. An analog controller is difficult to modify or redesign once
implemented in hardware. A digital controller is implemented in firmware or
software and its modification is possible without a complete replacement of
the original controller. Furthermore, the structure of the digital controller need
not follow one of the simple forms that are typically used in analog control.
More complex controller structures involve a few extra arithmetic operations
and are easily realizable.
Speed. The speed of computer hardware has increased exponentially since the
1980s. This increase in processing speed has made it possible to sample and
process control signals at very high speeds. Because the interval between
samples, the sampling period, can be made very small, digital controllers
achieve performance that is essentially the same as that based on continuous
monitoring of the controlled variable.
Cost. Although the prices of most goods and services have steadily increased,
the cost of digital circuitry continues to decrease. Advances in very large-scale
integration (VLSI) technology have made it possible to manufacture better,
faster, and more reliable integrated circuits and to offer them to the consumer
at a lower price. This has made the use of digital controllers more economical
even for small, low-cost applications.
1.2 The structure of a digital control system
To control a physical system or process using a digital controller, the controller
must receive measurements from the system, process them, and then send control
signals to the actuator that effects the control action. In almost all applications,
both the plant and the actuator are analog systems. This is a situation where the
controller and the controlled do not “speak the same language,” and some form
of translation is required. The translation from controller language (digital) to
1.3 Examples of digital control systems
Reference
Input
Computer
DAC
Actuator
and Process
ADC
Sensor
Controlled
Variable
FIGURE 1.1
Configuration of a digital control system.
physical process language (analog) is performed by a digital-to-analog converter,
or DAC. The translation from process language to digital controller language
is performed by an analog-to-digital converter, or ADC. A sensor is needed to
monitor the controlled variable for feedback control. The combination of the
elements discussed here in a control loop is shown in Figure 1.1. Variations on
this control configuration are possible. For example, the system could have
several reference inputs and controlled variables, each with a loop similar to
that of Figure 1.1. The system could also include an inner loop with digital
or analog control.
1.3 Examples of digital control systems
In this section, we briefly discuss examples of control systems where digital
implementation is now the norm. There are many other examples of industrial
processes that are digitally controlled, and the reader is encouraged to seek other
examples from the literature.
1.3.1 Closed-loop drug delivery system
Several chronic diseases require the regulation of the patient’s blood levels of a
specific drug or hormone. For example, some diseases involve the failure of the
body’s natural closed-loop control of blood levels of nutrients. Most prominent
among these is the disease diabetes, where the production of the hormone insulin
that controls blood glucose levels is impaired.
To design a closed-loop drug delivery system, a sensor is utilized to measure
the levels of the regulated drug or nutrient in the blood. This measurement is converted to digital form and fed to the control computer, which drives a pump that
injects the drug into the patient’s blood. A block diagram of the drug delivery
system is shown in Figure 1.2. See Carson and Deutsch (1992) for a more detailed
example of a drug delivery system.
3
4
CHAPTER 1 Introduction to Digital Control
Drug Tank
Computer
Regulated
Drug
or Nutrient
Drug
Pump
Blood
Sensor
(a)
Reference
Blood
Level
Computer
DAC
Drug
Pump
ADC
Blood
Sensor
Patient
Regulated
Drug
or Nutrient
(b)
FIGURE 1.2
Drug delivery digital control system. (a) Schematic of a drug delivery system.
(b) Block diagram of a drug delivery system.
1.3.2 Computer control of an aircraft turbojet engine
To achieve the high performance required for today’s aircraft, turbojet engines
employ sophisticated computer control strategies. A simplified block diagram for
turbojet computer control is shown in Figure 1.3. The control requires feedback
of the engine state (speed, temperature, and pressure), measurements of the aircraft
state (speed and direction), and pilot command.
1.3.3 Control of a robotic manipulator
Robotic manipulators are capable of performing repetitive tasks at speeds and
accuracies that far exceed those of human operators. They are now widely used
in manufacturing processes such as spot welding and painting. To perform their
tasks accurately and reliably, manipulator hand (or end-effector) positions and
velocities are controlled digitally. Each motion or degree of freedom (D.O.F.)
1.3 Examples of digital control systems
(a)
Pilot
Command
Computer
DAC
Turbojet
Engine
ADC
Engine
Sensors
Aircraft
Aircraft
State
Engine
State
Aircraft
Sensors
ADC
(b)
FIGURE 1.3
Turbojet engine control system. (a) F-22 military fighter aircraft. (b) Block diagram of an
engine control system.
of the manipulator is positioned using a separate position control system. All the
motions are coordinated by a supervisory computer to achieve the desired speed
and positioning of the end-effector. The computer also provides an interface
between the robot and the operator that allows programming the lower-level controllers and directing their actions. The control algorithms are downloaded from
the supervisory computer to the control computers, which are typically specialized
microprocessors known as digital signal processing (DSP) chips. The DSP chips
execute the control algorithms and provide closed-loop control for the manipulator. A simple robotic manipulator is shown in Figure 1.4a, and a block diagram
of its digital control system is shown in Figure 1.4b. For simplicity, only one
5
6
CHAPTER 1 Introduction to Digital Control
(a)
Reference
Trajectory Supervisory
Computer
Computers
DAC
Manipulator
ADC
Position
Sensors
ADC
Velocity
Sensors
(b)
FIGURE 1.4
Robotic manipulator control system. (a) 3-D.O.F. robotic manipulator. (b) Block diagram
of a manipulator control system.
motion control loop is shown in Figure 1.4, but there are actually n loops for an
n-D.O.F. manipulator.
Resources
Carson, E.R., Deutsch, T., 1992. A spectrum of approaches for controlling diabetes.
Control Syst. Mag. 12 (6), 2531.
Chen, C.T., 1993. Analog and Digital Control System Design. SaundersHBJ.
Koivo, A.J., 1989. Fundamentals for Control of Robotic Manipulators. Wiley.
Shaffer, P.L., 1990. A multiprocessor implementation of a real-time control of turbojet
engine. Control Syst. Mag. 10 (4), 3842.
Problems
PROBLEMS
1.1
A fluid level control system includes a tank, a level sensor, a fluid source,
and an actuator to control fluid inflow. Consult any classical control text1 to
obtain a block diagram of an analog fluid control system. Modify the block
diagram to show how the fluid level could be digitally controlled.
1.2
If the temperature of the fluid in Problem 1.1 is to be regulated together
with its level, modify the analog control system to achieve the additional
control. (Hint: An additional actuator and sensor are needed.) Obtain a
block diagram for the two-input-two-output control system with digital
control.
1.3
Position control servos are discussed extensively in classical control texts.
Draw a block diagram for a direct current motor position control system
after consulting your classical control text. Modify the block diagram to
obtain a digital position control servo.
1.4
Repeat Problem 1.3 for a velocity control servo.
1.5
A ballistic missile (see Figure P1.5) is required to follow a predetermined
flight path by adjusting its angle of attack α (the angle between its axis
and its velocity vector v). The angle of attack is controlled by adjusting the
thrust angle δ (angle between the thrust direction and the axis of the missile).
Draw a block diagram for a digital control system for the angle of attack,
including a gyroscope to measure the angle α and a motor to adjust the thrust
angle δ.
α
Velocity
Vector v
Thrust
Direction
δ
FIGURE P1.5
Missile angle-of-attack control.
1.6
1
A system is proposed to remotely control a missile from an earth station.
Because of cost and technical constraints, the missile coordinates would be
measured every 20 seconds for a missile speed of up to 0.5 mm/s. Is such a
control scheme feasible? What would the designers need to do to eliminate
potential problems?
See, for example, Van deVegte, J., 1994. Feedback Control Systems, Prentice Hall.
7
8
CHAPTER 1 Introduction to Digital Control
1.7
The control of the recording head of a dual actuator hard disk drive (HDD)
requires two types of actuators to achieve the required high real density.
The first is a coarse voice coil motor (VCM) with a large stroke but slow
dynamics, and the second is a fine piezoelectric transducer (PZT) with a
small stroke and fast dynamics. A sensor measures the head position, and
the position error is fed to a separate controller for each actuator. Draw a
block diagram for a dual actuator digital control system for the HDD.2
1.8
In a planar contour tracking task performed by a robot manipulator, the
robot end-effector is required to track the contour of an unknown object
with a given reference tangential velocity and by applying a given force to
the object in the normal direction. For this purpose a force sensor can be
applied on the end-effector, while the end-effector velocity can be determined by means of the joint velocities. Draw a block diagram of the digital
control system.3
1.9
A typical main irrigation canal consists of several pools separated by gates
that are used for regulating the water distribution from one pool to the next.
In automatically regulated canals, the controlled variables are the water
levels, the manipulated variables are the gate positions, and the fundamental
perturbation variables are the unknown offtake discharges.4 Draw a block
diagram of the control scheme.
2
Ding, J., Marcassa, F., Wu, S.-C., Tomizuka, M., 2006. Multirate control for computational saving,
IEEE Trans. Control Systems Tech. 14 (1), 165169.
3
Jatta, F., Legnani, G., Visioli, A., Ziliani, G., 2006. On the use of velocity feedback in hybrid
force/velocity control of industrial manipulators, Control Engineering Practice 14, 10451055.
4
Feliu-Battle, V., Rivas Perez, R., Sanchez Rodriguez, L., 2007. Fractional robust control of main
irrigation canals with variable dynamic parameters, Control Engineering Practice 15, 673686.
CHAPTER
Discrete-Time Systems
2
OBJECTIVES
After completing this chapter, the reader will be able to do the following:
1. Explain why difference equations result from digital control of analog systems.
2. Obtain the z-transform of a given time sequence and the time sequence
corresponding to a function of z.
3. Solve linear time-invariant (LTI) difference equations using the z-transform.
4. Obtain the z-transfer function of an LTI system.
5. Obtain the time response of an LTI system using its transfer function or impulse
response sequence.
6. Obtain the modified z-transform for a sampled time function.
7. Select a suitable sampling period for a given LTI system based on its dynamics.
Digital control involves systems whose control is updated at discrete time instants.
Discrete-time models provide mathematical relations between the system variables
at these time instants. In this chapter, we develop the mathematical properties of
discrete-time models that are used throughout the remainder of the text. For most
readers, this material provides a concise review of material covered in basic courses
on control and system theory. However, the material is self-contained, and familiarity with discrete-time systems is not required. We begin with an example that
illustrates how discrete-time models arise from analog systems under digital
control.
2.1 Analog systems with piecewise constant inputs
In most engineering applications, it is necessary to control a physical system
or plant so that it behaves according to given design specifications. Typically, the
plant is analog, the control is piecewise constant, and the control action is updated
periodically. This arrangement results in an overall system that is conveniently
described by a discrete-time model. We demonstrate this concept using a simple
example.
9
10
CHAPTER 2 Discrete-Time Systems
EXAMPLE 2.1
Consider the tank control system in Figure 2.1. In the figure, lowercase letters denote
perturbations from fixed steady-state values. The variables are defined as
•
•
•
•
•
H 5 steady-state fluid height in the tank
h 5 height perturbation from the nominal value
Q 5 steady-state flow rate through the tank
qi 5 inflow perturbation from the nominal value
q0 5 outflow perturbation from the nominal value
It is necessary to maintain a constant fluid level H by adjusting the fluid flow rate into
the tank around Q. Obtain an analog mathematical model of the tank, and use it to obtain
a discrete-time model for the system with piecewise constant inflow perturbation qi and
output h.
Solution
Although the fluid system is nonlinear, a linear model can satisfactorily describe the system
under the assumption that fluid level is regulated around a constant value. The linearized
model for the outflow valve is analogous to an electrical resistor and is given by
h 5 R q0
where h is the perturbation in tank level from nominal, q0 is the perturbation in the outflow
from the tank from a nominal level Q, and R is the fluid resistance of the valve.
Assuming an incompressible fluid, the principle of conservation of mass reduces to
the volumetric balance: rate of fluid volume increase 5 rate of fluid volume in 2 rate of
fluid volume out:
dCðh 1 HÞ
5 ðqi 1 QÞ 2 ðqo 1 QÞ
dt
where C is the area of the tank or its fluid capacitance. The term H is a constant and its
derivative is zero, and the term Q cancels so that the remaining terms only involve perturbations. Substituting for the outflow q0 from the linearized valve equation into the volumetric
fluid balance gives the analog mathematical model
dh
h qi
1 5
C
dt
τ
qi
h
H
qo
FIGURE 2.1
Fluid level control system.
2.2 Difference equations
where τ 5 RC is the fluid time constant for the tank. The solution of this differential
equation is
ð
1 t 2ðt2λÞ=τ
hðtÞ 5 e2ðt2t0 Þ=τ hðt0 Þ 1
e
qi ðλÞdλ
C t0
Let qi be constant over each sampling period T—that is, qi(t) 5 qi(k) 5 constant for t in
the interval [k T, (k 1 1) T]. Then we can solve the analog equation over any sampling
period to obtain
hðk 1 1Þ 5 e2T=τ hðkÞ 1 R½1 2 e2T=τ qi ðkÞ;
k 5 0; 1; 2; . . .
where the variables at time kT are denoted by the argument k. This is the desired discretetime model describing the system with piecewise constant control. Details of the solution
are left as an exercise (Problem 2.1).
The discrete-time model obtained in Example 2.1 is known as a difference equation.
Because the model involves a linear time-invariant analog plant, the equation is linear
time invariant. Next, we briefly discuss difference equations, and then we introduce a
transform used to solve them.
2.2 Difference equations
Difference equations arise in problems where the independent variable, usually
time, is assumed to have a discrete set of possible values. The nonlinear difference equation
yðk 1 nÞ 5 f ½yðk 1 n 2 1Þ; yðk 1 n 2 2Þ; . . . ; yðk 1 1Þ; yðkÞ; uðk 1 nÞ;
uðk 1 n 2 1Þ; . . . ; uðk 1 1Þ; uðkÞ
(2.1)
with forcing function u(k) is said to be of order n because the difference between
the highest and lowest time arguments of y(.) and u(.) is n. The equations we deal
with in this text are almost exclusively linear and are of the form
yðk 1 nÞ 1 an21 yðk 1 n 2 1Þ 1 ? 1 a1 yðk 1 1Þ 1 a0 yðkÞ
5 bn uðk 1 nÞ 1 bn21 uðk 1 n 2 1Þ 1 ? 1 b1 uðk 1 1Þ 1 b0 uðkÞ
(2.2)
We further assume that the coefficients ai, bi, i 5 0, 1, 2, . . . , are constant. The
difference equation is then referred to as linear time invariant, or LTI. If the forcing
function u(k) is equal to zero, the equation is said to be homogeneous.
EXAMPLE 2.2
For each of the following difference equations, determine the order of the equation. Is the
equation (a) linear, (b) time invariant, or (c) homogeneous?
1. y(k 1 2) 1 0.8y(k 1 1) 1 0.07y(k)u(k)
2. y(k 1 4) 1 sin(0.4k)y(k 1 1) 1 0.3y(k) 5 0
3. y(k 1 1) 5 20.1y2(k)
11
12
CHAPTER 2 Discrete-Time Systems
Solution
1. The equation is second order. All terms enter the equation linearly and have constant
coefficients. The equation is therefore LTI. A forcing function appears in the equation,
so it is nonhomogeneous.
2. The equation is fourth order. The second coefficient is time dependent, but all the
terms are linear and there is no forcing function. The equation is therefore linear time
varying and homogeneous.
3. The equation is first order. The right-hand side (RHS) is a nonlinear function of y(k),
but does not include a forcing function or terms that depend on time explicitly. The
equation is therefore nonlinear, time invariant, and homogeneous.
Difference equations can be solved using classical methods analogous to those
available for differential equations. Alternatively, z-transforms provide a convenient
approach for solving LTI equations, as discussed in the next section.
2.3 The z-transform
The z-transform is an important tool in the analysis and design of discrete-time
systems. It simplifies the solution of discrete-time problems by converting LTI
difference equations to algebraic equations and convolution to multiplication. Thus,
it plays a role similar to that served by Laplace transforms in continuous-time problems. Because we are primarily interested in application to digital control systems,
this brief introduction to the z-transform is restricted to causal signals (i.e., signals
with zero values for negative time) and the one-sided z-transform.
The following are two alternative definitions of the z-transform.
DEFINITION 2.1
Given the causal sequence {u0, u1, u2, . . . , uk, . . .}, its z-transform is defined as
UðzÞ 5 u0 1 u1 z21 1 u2 z22 1 ? 1 uk z2k
N
X
5
uk z2k
(2.3)
k50
The variable z21 in the preceding equation can be regarded as a time delay
operator. The z-transform of a given sequence can be easily obtained as in the
following example.
DEFINITION 2.2
Given the impulse train representation of a discrete-time signal,
u ðtÞ 5 u0 δðtÞ 1 u1 δðt 2 TÞ 1 u2 δðt 2 2TÞ 1 ? 1 uk δðt 2 kTÞ 1 . . .
N
X
uk δðt 2 kTÞ
5
k50
(2.4)
2.3 The z-transform
the Laplace transform of (2.4) is
U ðsÞ 5 u0 1 u1 e2sT 1 u2 e22sT 1 ? 1 uk e2ksT 1 . . .
N
X
uk ðe2sT Þk
5
(2.5)
k50
Let z be defined by
z 5 esT
(2.6)
Then, substituting from (2.6) in (2.5) yields the z-transform expression (2.3).
EXAMPLE 2.3
Obtain the z-transform of the sequence fuk gN
k50 5 f1; 3; 2; 0; 4; 0; 0; 0; . . .g.
Solution
Applying Definition 2.1 gives U(z) 5 1 1 3z21 1 2z22 1 4z24.
Although the preceding two definitions yield the same transform, each has its
advantages and disadvantages. The first definition allows us to avoid the use of
impulses and the Laplace transform. The second allows us to treat z as a complex
variable and to use some of the familiar properties of the Laplace transform, such
as linearity.
Clearly, it is possible to use Laplace transformation to study discrete time,
continuous time, and mixed systems. However, the z-transform offers significant
simplification in notation for discrete-time systems and greatly simplifies their
analysis and design.
2.3.1 z-Transforms of standard discrete-time signals
Having defined the z-transform, we now obtain the z-transforms of commonly
used discrete-time signals such as the sampled step, exponential, and the discretetime impulse. The following identities are used repeatedly to derive several
important results:
n
X
12 an11
;
a 6¼ 1
ak 5
12 a
k50
N
X
1
;
jaj , 1
ak 5
12 a
k50
EXAMPLE 2.4: UNIT IMPULSE
Consider the discrete-time impulse (Figure 2.2)
1; k 5 0
uðkÞ 5 δðkÞ 5
0; k ¼
6 0
(2.7)
13
14
CHAPTER 2 Discrete-Time Systems
1
k
1
0
–1
FIGURE 2.2
Discrete-time impulse.
Applying Definition 2.1 gives the z-transform
UðzÞ 5 1
Alternatively, one may consider the impulse-sampled version of the delta function
u (t) 5 δ(t). This has the Laplace transform
U ðsÞ 5 1
Substitution from (2.6) has no effect. Thus, the z-transform obtained using Definition 2.2
is identical to that obtained using Definition 2.1.
EXAMPLE 2.5: SAMPLED STEP
Consider the sequence fuk gN
k50 5 f1; 1; 1; 1; 1; 1; . . .g of Figure 2.3. Definition 2.1 gives
the z-transform
UðzÞ 5 1 1 z21 1 z22 1 z23 1 ? 1 z2k 1 . . .
N
X
5
z2k
k50
Using the identity (2.7) gives the following closed-form expression for the z-transform:
1
1 2 z21
z
5
z21
UðzÞ 5
Note that (2.7) is only valid for jzj , 1. This implies that the z-transform expression we
obtain has a region of convergence outside which is not valid. The region of convergence
1
…
k
–1
FIGURE 2.3
Sampled unit step.
0
1
2
3
2.3 The z-transform
must be clearly given when using the more general two-sided transform with functions that
are nonzero for negative time. However, for the one-sided z-transform and time functions
that are zero for negative time, we can essentially extend regions of convergence and use
the z-transform in the entire z-plane.1
1
The idea of extending the definition of a complex function to the entire complex plane is known as analytic
continuation. For a discussion of this topic, consult any text on complex analysis.
EXAMPLE 2.6: EXPONENTIAL
Let
uðkÞ 5
ak ;
0;
k$0
k,0
Figure 2.4 shows the case where 0 , a , 1. Definition 2.1 gives the z-transform
UðzÞ 5 1 1 az21 1 a2 z22 1 ? 1 ak z2k 1 . . .
Using (2.7), we obtain
1
12 ða=zÞ
z
5
z2a
UðzÞ 5
As in Example 2.5, we can use the transform in the entire z-plane in spite of the validity
condition for (2.7) because our time function is zero for negative time.
1
…
a
a2
–1
0
1
a3
2
k
3
FIGURE 2.4
Sampled exponential.
2.3.2 Properties of the z-transform
The z-transform can be derived from the Laplace transform as shown in
Definition 2.2. Hence, it shares several useful properties with the Laplace transform, which can be stated without proof. These properties can also be easily
proved directly, and the proofs are left as an exercise for the reader. Proofs are
provided for properties that do not obviously follow from the Laplace transform.
15
16
CHAPTER 2 Discrete-Time Systems
Linearity
This equation follows directly from the linearity of the Laplace transform.
Zfαf1 ðkÞ 1 β f2 ðkÞg 5 αF1 ðzÞ 1 β F2 ðzÞ
(2.8)
EXAMPLE 2.7
Find the z-transform of the causal sequence
f ðkÞ 5 2 3 1ðkÞ 1 4δðkÞ;
k 5 0; 1; 2; . . .
Solution
Using linearity, the transform of the sequence is
FðzÞ 5 Zf2 3 1ðkÞ 1 4δðkÞg 5 2 Zf1ðkÞg 1 4 ZfδðkÞg 5
2z
6z 2 4
145
z21
z21
Time delay
This equation follows from the time delay property of the Laplace transform and
equation (2.6).
Zff ðk 2 nÞg 5 z2n FðzÞ
(2.9)
EXAMPLE 2.8
Find the z-transform of the causal sequence
f ðkÞ 5
4;
0;
k 5 2; 3; . . .
otherwise
Solution
The given sequence is a sampled step starting at k 5 2 rather than k 5 0 (i.e., it is delayed
by two sampling periods). Using the delay property, we have
FðzÞ 5 Zf4 3 1ðk 2 2Þg 5 4 z22 Zf1ðkÞg 5 z22
4z
4
5
z 2 1 zðz 2 1Þ
Time advance
Zff ðk 1 1Þg 5 zFðzÞ 2 zf ð0Þ
Zff ðk 1 nÞg 5 zn FðzÞ 2 zn f ð0Þ 2 zn21 f ð1Þ 2 ? 2 zf ðn 2 1Þ
(2.10)
2.3 The z-transform
PROOF
Only the first part of the theorem is proved here. The second part can be easily proved by
induction. We begin by applying the z-transform Definition 2.1 to a discrete-time function
advanced by one sampling interval. This gives
Zff ðk 1 1Þg 5
N
X
f ðk 1 1Þz2k
k50
N
X
f ðk 1 1Þz2ðk11Þ
5z
k50
Now add and subtract the initial condition f(0) to obtain
("
#
)
N
X
f ðk 1 1Þz2ðk11Þ 2 f ð0Þ
Zff ðk 1 1Þg 5 z f ð0Þ 1
k50
Next, change the index of summation to m 5 k 1 1 and rewrite the z-transform as
("
#
)
N
X
f ðmÞz2m 2 f ð0Þ
Zff ðk 1 1Þg 5 z
m50
5 zFðzÞ 2 zf ð0Þ
EXAMPLE 2.9
Using the time advance property, find the z-transform of the causal sequence
ff ðkÞg 5 f4; 8; 16; . . .g
Solution
The sequence can be written as
f ðkÞ 5 2k12 5 gðk 1 2Þ;
k 5 0; 1; 2; . . .
where g(k) is the exponential time function
gðkÞ 5 2k ;
k 5 0; 1; 2; . . .
Using the time advance property, we write the transform
FðzÞ 5 z2 GðzÞ 2 z2 gð0Þ 2 zgð1Þ 5 z2
z
4z
2 z2 2 2z 5
z22
z22
Clearly, the solution can be obtained directly by rewriting the sequence as
ff ðkÞg 5 4f1; 2; 4; . . .g
and using the linearity of the z-transform.
Multiplication by exponential
Zfa2k f ðkÞg 5 FðazÞ
(2.11)
17
18
CHAPTER 2 Discrete-Time Systems
PROOF
LHS 5
N
X
a2k f ðkÞz2k 5
k50
N
X
f ðkÞðazÞ2k 5 FðazÞ
k50
EXAMPLE 2.10
Find the z-transform of the exponential sequence
f ðkÞ 5 e2αkT ;
k 5 0; 1; 2; . . .
Solution
Recall that the z-transform of a sampled step is
FðzÞ 5 ð12 z21 Þ21
and observe that f(k) can be rewritten as
f ðkÞ 5 ðeαT Þ2k 3 1;
k 5 0; 1; 2; . . .
Then apply the multiplication by exponential property to obtain
Z ðeαT Þ2k f ðkÞ 5 ½12ðeαT zÞ21 21 5
z
z 2 e2αT
This is the same as the answer obtained in Example 2.6.
Complex differentiation
d m
Z k f ðkÞ 5 2z
FðzÞ
dz
m
(2.12)
PROOF
To prove the property by induction, we first establish its validity for m 5 1. Then we assume
its validity for any m and prove it for m 1 1. This establishes its validity for 1 1 1 5 2,
then 2 1 1 5 3, and so on.
For m 5 1, we have
N
N
X
X
d 2k
k f ðkÞz2k 5
f ðkÞ 2z
z
dz
k50
k50
N
d X
d
5 2z
f ðkÞz2k 5 2z
FðzÞ
dz k50
dz
Zfk f ðkÞg 5
Next, let the statement be true for any m and define the sequence
fm ðkÞ 5 km f ðkÞ;
k 5 0; 1; 2; . . .
2.3 The z-transform
and obtain the transform
N
X
k fm ðkÞz2k
k50
N
X
d 2k
5
fm ðkÞ 2z
z
dz
k50
X
d N
d
5 2z
fm ðkÞz2k 5 2z
Fm ðzÞ
dz k50
dz
Zfk fm ðkÞg 5
Substituting for Fm(z), we obtain the result
d m11
FðzÞ
Z km11 f ðkÞ 5 Z k fm ðkÞ 5 2z
dz
EXAMPLE 2.11
Find the z-transform of the sampled ramp sequence
f ðkÞ 5 k;
k 5 0; 1; 2; . . .
Solution
Recall that the z-transform of a sampled step is
FðzÞ 5
z
z21
and observe that f(k) can be rewritten as
f ðkÞ 5 k 3 1;
k 5 0; 1; 2; . . .
Then apply the complex differentiation property to obtain
d
z
ðz 2 1Þ 2 z
z
5
5 ð2zÞ
Z fk 3 1g 5 2z
dz z 21
ðz21Þ2
ðz21Þ2
2.3.3 Inversion of the z-transform
Because the purpose of z-transformation is often to simplify the solution of time
domain problems, it is essential to inverse-transform z-domain functions. As in the
case of Laplace transforms, a complex integral can be used for inverse transformation. This integral is difficult to use and is rarely needed in engineering applications.
Two simpler approaches for inverse z-transformation are discussed in this section.
Long division
This approach is based on Definition 2.1, which relates a time sequence to its
z-transform directly. We first use long division to obtain as many terms as desired
of the z-transform expansion; then we use the coefficients of the expansion to
19
20
CHAPTER 2 Discrete-Time Systems
write the time sequence. The following two steps give the inverse z-transform of
a function F(z):
1. Using long division, expand F(z) as a series to obtain
Ft ðzÞ 5 f0 1 f1 z21 1 ? 1 fi z2i 5
i
X
fk z2k
k50
2. Write the inverse transform as the sequence
ff0 ; f1 ; . . .; fi ; . . .g
The number of terms i obtained by long division is selected to yield a sufficient
number of points in the time sequence.
EXAMPLE 2.12
Obtain the inverse z-transform of the function
FðzÞ 5
z11
z2 1 0:2z 1 0:1
Solution
1. Long Division
z −1 + 0 .8 z −2 − 0 .26 z −3 + . . . . . .
z + 0 .2 z + 0 .1 z + 1
2
)
z + 0 .2 + 0 .1z −1
0 .8 − 0 .10 z −1
0 .8 + 0 .1 6 z −1 + 0 .08 z −2
−0 .26 z −1 − . . .
Thus, Ft(z) 5 0 1 z21 1 0.8z22 2 0.26z23
2. Inverse Transformation
ffk g 5 f0; 1; 0:8; 20:26; . . .g
Partial fraction expansion
This method is almost identical to that used in inverting Laplace transforms.
However, because most z-functions have the term z in their numerator, it is often
convenient to expand F(z)/z rather than F(z). As with Laplace transforms, partial
fraction expansion allows us to write the function as the sum of simpler functions
2.3 The z-transform
that are the z-transforms of known discrete-time functions. The time functions are
available in z-transform tables such as the table provided in Appendix I.
The procedure for inverse z-transformation is
1. Find the partial fraction expansion of F(z)/z or F(z).
2. Obtain the inverse transform f(k) using the z-transform tables.
We consider three types of z-domain functions F(z): functions with simple
(nonrepeated) real poles, functions with complex conjugate and real poles, and
functions with repeated poles. We discuss examples that demonstrate partial fraction expansion and inverse z-transformation in each case.
CASE 1 SIMPLE REAL ROOTS
The most convenient method to obtain the partial fraction expansion of a function with
simple real roots is the method of residues. The residue of a complex function F(z) at
a simple pole zi is given by
Ai 5 ðz 2 zi ÞFðzÞz-zi
(2.13)
This is the partial fraction coefficient of the ith term of the expansion
FðzÞ 5
n
X
Ai
z
2
zi
i51
(2.14)
Because most terms in the z-transform tables include a z in the numerator (see
Appendix I), it is often convenient to expand F(z)/z and then to multiply both sides
by z to obtain an expansion whose terms have a z in the numerator. Except for functions
that already have a z in the numerator, this approach is slightly longer but has the
advantage of simplifying inverse transformation. Both methods are examined through
Example 2.13.
EXAMPLE 2.13
Obtain the inverse z-transform of the function
FðzÞ 5
z11
z2 1 0:3z 1 0:02
Solution
It is instructive to solve this problem using two different methods. First, we divide by z;
then we obtain the partial fraction expansion.
1. Partial Fraction Expansion
Dividing the function by z, we expand as
FðzÞ
z11
5 2
z
zðz 1 0:3z 1 0:02Þ
5
A
B
C
1
1
z
z 1 0:1 z 1 0:2
21
22
CHAPTER 2 Discrete-Time Systems
where the partial fraction coefficients are given by
FðzÞ
1
5 Fð0Þ 5
A5z
5 50
z z50
0:02
B 5 ðz 1 0:1Þ
C 5 ðz 1 0:2Þ
FðzÞ
z
FðzÞ
z
5
12 0:1
5 290
ð20:1Þð0:1Þ
5
12 0:2
5 40
ð20:2Þð20:1Þ
z520:1
z520:2
Thus, the partial fraction expansion is
FðzÞ 5
50z
90z
40z
2
1
z
z 1 0:1 z 1 0:2
2. Table Lookup
f ðkÞ 5
50δðkÞ 2 90ð20:1Þk 1 40ð20:2Þk ;
0;
k$0
k,0
Note that f(0) 5 0, so the time sequence can be rewritten as
290ð20:1Þk 1 40ð20:2Þk ; k $ 1
f ðkÞ 5
0;
k,1
Now, we solve the same problem without dividing by z.
1. Partial Fraction Expansion
We obtain the partial fraction expansion directly
FðzÞ 5
5
z11
z2 1 0:3z 1 0:02
A
B
1
z 1 0:1 z 1 0:2
where the partial fraction coefficients are given by
12 0:1
59
0:1
12 0:2
5 28
B 5 ðz 1 0:2ÞFðzÞz520:2 5
20:1
A 5 ðz 1 0:1ÞFðzÞz520:1 5
Thus, the partial fraction expansion is
FðzÞ 5
9
8
2
z 1 0:1 z 1 0:2
2. Table Lookup
Standard z-transform tables do not include the terms in the expansion of F(z). However,
F(z) can be written as
FðzÞ 5
9z 21
8z 21
z 2
z
z 1 0:1
z 1 0:2
2.3 The z-transform
Then we use the delay theorem to obtain the inverse transform
f ðkÞ 5
9ð20:1Þk21 2 8ð20:2Þk21 ; k $ 1
0;
k,1
Verify that this is the answer obtained earlier when dividing by z written in a different form
(observe the exponent in the preceding expression).
Although it is clearly easier to obtain the partial fraction expansion without
dividing by z, inverse transforming requires some experience. There are situations where division by z may actually simplify the calculations, as seen in
Example 2.14.
EXAMPLE 2.14
Find the inverse z-transform of the function
FðzÞ 5
z
ðz 1 0:1Þðz 1 0:2Þðz 1 0:3Þ
Solution
1. Partial Fraction Expansion
Dividing by z simplifies the numerator and gives the expansion
FðzÞ
1
5
z
ðz 1 0:1Þðz 1 0:2Þðz 1 0:3Þ
5
A
B
C
1
1
z 1 0:1 z 1 0:2 z 1 0:3
where the partial fraction coefficients are
A 5 ðz 1 0:1Þ
B 5 ðz 1 0:2Þ
FðzÞ
z
FðzÞ
z
FðzÞ
C 5 ðz 1 0:3Þ
z
5
1
5 50
ð0:1Þð0:2Þ
5
1
5 2100
ð20:1Þð0:1Þ
5
1
5 50
ð20:2Þð20:1Þ
z520:1
z520:2
z520:3
Thus, the partial fraction expansion after multiplying by z is
FðzÞ 5
50z
100z
50z
2
1
z 1 0:1 z 1 0:2 z 1 0:3
2. Table Lookup
f ðkÞ 5
50ð20:1Þk 2 100ð20:2Þk 1 50ð20:3Þk ;
0;
k$0
k,0
23
24
CHAPTER 2 Discrete-Time Systems
CASE 2 COMPLEX CONJUGATE AND SIMPLE REAL ROOTS
For a function F(z) with real and complex poles, the partial fraction expansion includes
terms with real roots and others with complex roots. Assuming that F(z) has real
coefficients, then its complex roots occur in complex conjugate pairs and can be combined
to yield a function with real coefficients and a quadratic denominator. To inverse-transform
such a function, use the following z-transforms (see Appendix I):
Zfe2αk sinðkωd Þg 5
Zfe2αk cosðkωd Þg 5
e2α sinðωd Þz
z2 2 2e2α cosðωd Þz 1 e22α
(2.15)
z½z 2 e2α cosðωd Þ
2 2e2α cosðωd Þz 1 e22α
(2.16)
z2
Note that in some tables the sampling period T is omitted, but then ωd is given in
radians and α is dimensionless. The denominators of the two transforms are identical and
have complex conjugate roots. The numerators can be scaled and combined to give the
desired inverse transform.
To obtain the partial fraction expansion, we use the residues method shown in Case 1.
With complex conjugate poles, we obtain the partial fraction expansion
FðzÞ 5
Az
A z
1
z 2 p z 2 p
(2.17)
We then inverse z-transform to obtain
f ðkÞ 5 Apk 1 A pk
5 jAjjpjk ½ejðθp k1θA Þ 1 e2jðθp k1θA Þ where θp and θA are the angle of the pole p and the angle of the partial fraction coefficient A,
respectively. We use the exponential expression for the cosine function to obtain
f ðkÞ 5 2jAjjpjk cosðθp k 1 θA Þ
(2.18)
Most modern calculators can perform complex arithmetic, and the residues
method is preferable in most cases. Alternatively, by equating coefficients, we
can avoid the use of complex arithmetic entirely, but the calculations can be quite
tedious. Example 2.15 demonstrates the two methods.
EXAMPLE 2.15
Find the inverse z-transform of the function
FðzÞ 5
z3 1 2z 1 1
ðz 2 0:1Þðz2 1 z 1 0:5Þ
Solution: Equating Coefficients
1. Partial Fraction Expansion
Dividing the function by z gives
FðzÞ
z3 1 2z 1 1
5
z
zðz 2 0:1Þðz2 1 z 1 0:5Þ
5
A1
A2
Az 1 B
1
1
z
z 2 0:1 z2 1 z 1 0:5
2.3 The z-transform
The first two coefficients can be easily evaluated as before. Thus,
A1 5 Fð0Þ 5 220
FðzÞ
D19:689
z
A2 5 ðz 2 0:1Þ
To evaluate the remaining coefficients, we multiply the equation by the denominator
and equate coefficients to obtain
z3 : A1 1 A2 1 A 5 1
z1 : 0:4 A1 1 0:5A2 2 0:1 B 5 2
where the coefficients of the third- and first-order terms yield separate equations in A and
B. Because A1 and A2 have already been evaluated, we can solve each of the two equations
for one of the remaining unknowns to obtain
AD1:311 BD 21:557
Had we chosen to equate coefficients without first evaluating A1 and A2, we would have
faced that considerably harder task of solving four equations in four unknowns. The
remaining coefficients can be used to check our calculations:
z0 : 20:05 A1 5 0:05ð20Þ 5 1
z2 : 0:9 A1 1 A2 2 0:1A 1 B 5 0:9ð220Þ 1 19:689 2 0:1ð1:311Þ 2 1:557D0
The results of these checks are approximate, because approximations were made in
the calculations of the coefficients. The partial fraction expansion is
FðzÞ 5 220 1
19:689 z 1:311 z2 2 1:557z
1
z 2 0:1
z2 1 z 1 0:5
2. Table Lookup
The first two terms of the partial fraction expansion can be easily found in the z-transform
tables. The third term resembles the transforms of a sinusoid multiplied by an exponential
if rewritten as
1:311 z2 2 1:557z
1:311z½z 2 e2α cosðωd Þ 2 Cze2α sinðωd Þ
5
2 2ð20:5Þz 1 0:5
z2 2 2e2α cosðωd Þz 1 e22α
z2
Starting with the constant term in the denominator, we equate coefficients to obtain
pffiffiffiffiffiffi
e2α 5 0:5 5 0:707
Next, the denominator z1 term gives
pffiffiffiffiffiffi
cosðωd Þ 5 20:5=e2α 5 2 0:5 5 20:707
Thus, ωd 5 3π/4, an angle in the second quadrant, with sin(ωd) 5 0.707.
Finally, we equate the coefficients of z1 in the numerator to obtain
21:311e2α cosðωd Þ 2 Ce2α sinðωd Þ 5 20:5ðC 2 1:311Þ 5 21:557
and solve for C 5 4.426. Referring to the z-transform tables, we obtain the inverse
transform
f ðkÞ 5 220δðkÞ 1 19:689ð0:1Þk 1 ð0:707Þk ½1:311cosð3πk=4Þ 2 4:426sinð3πk=4Þ
25
26
CHAPTER 2 Discrete-Time Systems
for positive time k. The sinusoidal terms can be combined using the trigonometric identities
sinðA 2 BÞ 5 sinðAÞcosðBÞ 2 sinðBÞcosðAÞ
sin21 ð1:311=4:616Þ 5 0:288
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
and the constant 4:616 5 ð1:311Þ2 1 ð4:426Þ2 . This gives
f ðkÞ 5 220δðkÞ 1 19:689ð0:1Þk 2 4:616ð0:707Þk sinð3πk=4 2 0:288Þ
Residues
1. Partial Fraction Expansion
Dividing by z gives
FðzÞ
z3 1 2z 1 1
5
z
zðz 2 0:1Þ½ðz10:5Þ2 1 0:52 5
A1
A2
A3
A3
1
1
1
z
z 2 0:1 z 1 0:5 2 j0:5 z 1 0:5 1 j0:5
The partial fraction expansion can be obtained as in the first approach
z3 12z11
D0:656 1 j2:213
A3 5
zðz20:1Þðz10:51j0:5Þ z520:51j0:5
FðzÞ 5 220 1
19:689 z ð0:656 1 j2:213Þz ð0:656 2 j2:213Þz
1
1
z 2 0:1
z 1 0:5 2 j0:5
z 1 0:5 1 j0:5
We convert the coefficient A3 from Cartesian to polar form:
A3 5 0:656 1 j2:213 5 2:308ej1:283
We inverse z-transform to obtain
f ðkÞ 5 220δðkÞ 1 19:689ð0:1Þk 1 4:616ð0:707Þk cosð3πk=4 1 1:283Þ
This is equal to the answer obtained earlier because 1.283 2 π/2 5 20.288.
CASE 3 REPEATED ROOTS
For a function F(z) with a repeated root of multiplicity r, r partial fraction coefficients are
associated with the repeated root. The partial fraction expansion is of the form
FðzÞ 5
NðzÞ
n
ðz2z1 Þr L z 2 zj
5
r
X
i51
n
X
A1i
Aj
1
r112i
z 2 zj
ðz2z1 Þ
j5r11
(2.19)
j5r11
The coefficients for repeated roots are governed by
1 di21
r
ðz2z
Þ
FðzÞ
;
A1;i 5
1
ði21Þ! dzi21
z-z1
i 5 1; 2; . . . ; r
The coefficients of the simple or complex conjugate roots can be obtained as before
using (2.13).
(2.20)
2.3 The z-transform
EXAMPLE 2.16
Obtain the inverse z-transform of the function
FðzÞ 5
1
z2 ðz 2 0:5Þ
Solution
1. Partial Fraction Expansion
Dividing by z gives
FðzÞ
1
A11
A12
A13
A4
5 3 1 2 1
1
5 3
z
z
z
z 2 0:5
z
z ðz 2 0:5Þ
where
A11 5 z3
FðzÞ 1 5
5 22
z z50 z20:5 z50
A12 5
1 d 3 FðzÞ d
1 21 5
5
5 24
z
2
1! dz
z z50 dz z 2 0:5 z50 ðz 2 0:5Þ z50
A13 5
1 d 2 3 FðzÞ z
2! dz2
z z50
5
1 d
21 1 ð21Þð22Þ 5
5 28
2
3
2 dz ðz 2 0:5Þ z50
2 ðz 2 0:5Þ z50
A4 5 ðz 2 0:5Þ
FðzÞ 1 5
58
z z50:5 z3 z50:5
Thus, we have the partial fraction expansion
FðzÞ 5
1
8z
5
2 2z22 2 4z21 2 8
z2 ðz 2 0:5Þ z 2 0:5
2. Table Lookup
The z-transform tables and Definition 2.1 yield
f ðkÞ 5
8ð0:5Þk 2 2δðk 2 2Þ 2 4δðk 2 1Þ 2 8δðkÞ; k $ 0
0;
k,0
Evaluating f(k) at k 5 0, 1, 2 yields
f ð0Þ 5 8 2 8 5 0
f ð1Þ 5 8ð0:5Þ 2 4 5 0
f ð2Þ 5 8ð0:5Þ2 2 2 5 0
We can therefore rewrite the inverse transform as
f ðkÞ 5
ð0:5Þk23 ;
0;
k$3
k,3
27
28
CHAPTER 2 Discrete-Time Systems
Note that the solution can be obtained directly using the delay theorem without the need
for partial fraction expansion because F(z) can be written as
FðzÞ 5
z
z23
z 2 0:5
The delay theorem and the inverse transform of an exponential yield the solution
obtained earlier.
2.3.4 The final value theorem
The final value theorem allows us to calculate the limit of a sequence as k tends
to infinity, if one exists, from the z-transform of the sequence. If one is only
interested in the final value of the sequence, this constitutes a significant shortcut.
The main pitfall of the theorem is that there are important cases where the limit
does not exist. The two main cases are as follows:
1. An unbounded sequence
2. An oscillatory sequence
The reader is cautioned against blindly using the final value theorem, because this
can yield misleading results.
THEOREM 2.1: THE FINAL VALUE THEOREM
If a sequence approaches a constant limit as k tends to infinity, then the limit is given by
z21
(2.21)
FðzÞ 5 Limðz 2 1ÞFðzÞ
f ðNÞ 5 Lim f ðkÞ 5 Lim
z-1
z-1
k-N
z
PROOF
Let f(k) have a constant limit as k tends to infinity; then the sequence can be expressed as
the sum
f ðkÞ 5 f ðNÞ 1 gðkÞ;
k 5 0; 1; 2; . . .
with g(k) a sequence that decays to zero as k tends to infinity; that is,
Lim f ðkÞ 5 f ðNÞ
k-N
The z-transform of the preceding expression is
FðzÞ 5
f ðNÞz
1 GðzÞ
z21
The final value f(N) is the partial fraction coefficient obtained by expanding F(z)/z as follows:
f ðNÞ 5 Limðz 2 1Þ
z-1
FðzÞ
5 Limðz 2 1ÞFðzÞ
z-1
z
2.4 Computer-aided design
EXAMPLE 2.17
Verify the final value theorem using the z-transform of a decaying exponential sequence
and its limit as k tends to infinity.
Solution
The z-transform pair of an exponential sequence is
e2akT
Z
!
z
z 2 e2aT
with a . 0. The limit as k tends to infinity in the time domain is
f ðNÞ 5 Lim e2akT 5 0
k-N
The final value theorem gives
z21
z
50
2aT
z-1
z
z2e
f ðNÞ 5 Lim
EXAMPLE 2.18
Obtain the final value for the sequence whose z-transform is
FðzÞ 5
z2 ðz 2 aÞ
ðz 2 1Þðz 2 bÞðz 2 cÞ
What can you conclude concerning the constants b and c if it is known that the limit
exists?
Solution
Applying the final value theorem, we have
f ðNÞ 5 Lim
zðz 2 aÞ
z-1 ðz 2 bÞðz 2 cÞ
5
12a
ð1 2 bÞð1 2 cÞ
To inverse z-transform the given function, one would have to obtain its partial fraction
expansion, which would include three terms: the transform of the sampled step, the transform of the exponential (b)k, and the transform of the exponential (c)k. Therefore, the
conditions for the sequence to converge to a constant limit and for the validity of the final
value theorem are jbj , 1 and jcj , 1.
2.4 Computer-aided design
In this text, we make extensive use of computer-aided design (CAD) and analysis
of control systems. We use MATLAB,2 a powerful package with numerous useful
commands. For the reader’s convenience, we list some MATLAB commands after
MATLABs is a copyright of MathWorks Inc., of Natick, Massachusetts.
2
29
30
CHAPTER 2 Discrete-Time Systems
covering the relevant theory. The reader is assumed to be familiar with the
CAD package but not with the digital system commands. We adopt the notation
of bolding all user commands throughout the text. Readers using other CAD
packages will find similar commands for digital control system analysis and
design.
MATLAB typically handles coefficients as vectors with the coefficients listed
in descending order. The function G(z) with numerator 5(z 1 3) and denominator
z3 1 0.1z2 1 0.4z is represented as the numerator polynomial
.. num 5 5 ½1; 3
and the denominator polynomial
.. den 5 ½1; 0:1; 0:4; 0
Multiplication of polynomials is equivalent to the convolution of their vectors of
coefficients and is performed using the command
.. denp 5 convðden1; den2Þ
where denp is the product of den1 and den2.
The partial fraction coefficients are obtained using the command
.. ½r;p;k 5 residueðnum; denÞ
where p represents the poles, r their residues, and k the coefficients of the polynomial resulting from dividing the numerator by the denominator. If the highest
power in the numerator is smaller than the highest power in the denominator, k is
zero. This is the usual case encountered in digital control problems.
MATLAB allows the user to sample a function and z-transform it with the
commands
.. g 5 tfðnum; denÞ
.. gd 5 c2dðg; 0:1; 'imp'Þ
Other useful MATLAB commands are available with the symbolic manipulation
toolbox.
ztrans z-transform
iztrans inverse z-transform
To use these commands, we must first define symbolic variables such as z, g, and
k with the command
.. syms z g k
Powerful commands for symbolic manipulations are also available through
packages such as MAPLE, MATHEMATICA, and MACSYMA.
2.5 z-Transform solution of difference equations
2.5 z-Transform solution of difference equations
By a process analogous to Laplace transform solution of differential equations,
one can easily solve linear difference equations. The equations are first transformed to the z-domain (i.e., both the right- and left-hand sides of the equation
are z-transformed). Then the variable of interest is solved for and inverse
z-transformed. To transform the difference equation, we typically use the time
delay or the time advance property. Inverse z-transformation is performed using
the methods of Section 2.3.
EXAMPLE 2.19
Solve the linear difference equation
xðk 1 2Þ 2 ð3=2Þxðk 1 1Þ 1 ð1=2ÞxðkÞ 5 1ðkÞ
with the initial conditions x(0) 5 1, x(1) 5 5/2.
Solution
1. z-transform
We begin by z-transforming the difference equation using (2.10) to obtain
½z2 XðzÞ 2 z2 xð0Þ 2 zxð1Þ 2ð3=2Þ½zXðzÞ 2 zxð0Þ 1ð1=2ÞXðzÞ 5 z=ðz 2 1Þ
2. Solve for X(z)
Then we substitute the initial conditions and rearrange terms to obtain
½z2 2ð3=2Þz 1ð1=2ÞXðzÞ 5 z=ðz 2 1Þ 1 z2 1ð5=2 2 3=2Þz
which we solve for
XðzÞ 5
z½1 1ðz 1 1Þðz 2 1Þ
z3
5
ðz 2 1Þðz 2 1Þðz 2 0:5Þ ðz21Þ2 ðz 2 0:5Þ
3. Partial Fraction Expansion
The partial fraction of X(z)/z is
XðzÞ
z2
A11
A12
A3
5
1
1
5
2
z 2 1 z 2 0:5
z
ðz21Þ ðz 2 0:5Þ ðz21Þ2
where
XðzÞ z2 1
5
5
52
z z51 z20:5 z51 1 2 0:5
XðzÞ z2 ð0:5Þ2
5
5
51
A3 5 ðz20:5Þ
z z50:5 ðz21Þ2 z50:5 ð0:521Þ2
A11 5 ðz21Þ2
To obtain the remaining coefficient, we multiply by the denominator and get the equation
z2 5 A11 ðz 2 0:5Þ 1 A12 ðz 2 0:5Þðz 2 1Þ 1 A3 ðz21Þ2
31
32
CHAPTER 2 Discrete-Time Systems
Equating the coefficient of z2 gives
z2 :1 5 A12 1 A3 5 A12 1 1
i:e:; A12 5 0
Thus, the partial fraction expansion in this special case includes two terms only. We
now have
XðzÞ 5
2z
z
1
z 2 0:5
ðz21Þ2
4. Inverse z-transformation
From the z-transform tables, the inverse z-transform of X(z) is
xðkÞ 5 2k 1 ð0:5Þk
2.6 The time response of a discrete-time system
The time response of a discrete-time linear system is the solution of the difference
equation governing the system. For the linear time-invariant (LTI) case, the
response due to the initial conditions and the response due to the input can be
obtained separately and then added to obtain the overall response of the system.
The response due to the input, or the forced response, is the convolution summation of its input and its response to a unit impulse. In this section, we derive this
result and examine its implications.
2.6.1 Convolution summation
The response of a discrete-time system to a unit impulse is known as the impulse
response sequence. The impulse response sequence can be used to represent the
response of a linear discrete-time system to an arbitrary input sequence
fuðkÞg 5 fuð0Þ; uð1Þ; . . .; uðiÞ; . . .g
(2.22)
To derive this relationship, we first represent the input sequence in terms of
discrete impulses as follows:
uðkÞ 5 uð0ÞδðkÞ 1 uð1Þδðk 2 1Þ 1 uð2Þδðk 2 2Þ 1 ? 1 uðiÞδðk 2 iÞ 1 . . .
N
X
uðiÞδðk 2 iÞ
5
(2.23)
i50
For a linear system, the principle of superposition applies, and the system output due to the input is the following sum of impulse response sequences:
fyðlÞg 5 fhðlÞguð0Þ 1 fhðl 2 1Þguð1Þ 1 fhðl 2 2Þguð2Þ 1 ? 1 fhðl 2 iÞguðiÞ 1 . . .
2.6 The time response of a discrete-time system
δ(k – i)
{h(k – i)}
…
iT
iT
LTI System
FIGURE 2.5
Response of a causal LTI discrete-time system to an impulse at iT.
Hence, the output at time k is given by
yðkÞ 5 hðkÞ uðkÞ 5
N
X
hðk 2 iÞuðiÞ
(2.24)
i50
where ( ) denotes the convolution operation.
For a causal system, the response due to an impulse at time i is an impulse
response starting at time i and the delayed response h(k 2 i) satisfies (Figure 2.5):
hðk 2 iÞ 5 0;
i.k
(2.25)
In other words, a causal system is one whose impulse response is a causal
time sequence. Thus, (2.24) reduces to
yðkÞ 5 uð0ÞhðkÞ 1 uð1Þhðk 2 1Þ 1 uð2Þhðk 2 2Þ 1 ? 1 uðkÞhð0Þ
k
X
uðiÞhðk 2 iÞ
5
(2.26)
i50
A simple change of summation variable (j 5 k 2 i) transforms (2.26) to
yðkÞ 5 uðkÞhð0Þ 1 uðk 2 1Þhð1Þ 1 uðk 2 2Þhð2Þ 1 ? 1 uð0ÞhðkÞ
k
X
uðk 2 jÞhðjÞ
5
(2.27)
j50
Equation (2.24) is the convolution summation for a noncausal system, whose
impulse response is nonzero for negative time, and it reduces to (2.26) for a
causal system. The summations for time-varying systems are similar, but the
impulse response at time i is h(k, i). Here, we restrict our analysis to LTI systems.
We can now summarize the result obtained in Theorem 2.2.
THEOREM 2.2: RESPONSE OF AN LTI SYSTEM
The response of an LTI discrete-time system to an arbitrary input sequence is given by the
convolution summation of the input sequence and the impulse response sequence of the
system.
33
34
CHAPTER 2 Discrete-Time Systems
Table 2.1 Input Components and Corresponding Output Components
Input
Response
Figure 2.6 Color
u(0) δ(k)
u(1) δ(k 2 1)
u(2) δ(k 2 2)
u(0) {h(k)}
u(1) {h(k 2 1)}
u(2) {h(k 2 2)}
White
Gray
Black
u(0) h(2)
k
u(1) h(1)
k
u(2) h(0)
k
1
0
2
3
FIGURE 2.6
Output at k 5 2.
To better understand the operations involved in convolution summation, we
evaluate one point in the output sequence using (2.24). For example,
yð2Þ 5
2
X
uðiÞhð2 2 iÞ
i50
5 uð0Þhð2Þ 1 uð1Þhð1Þ 1 uð2Þhð0Þ
From Table 2.1 and Figure 2.6, one can see the output corresponding to various components of the input of (2.23) and how they contribute to y(2). Note that
future input values do not contribute because the system is causal.
2.6.2 The convolution theorem
The convolution summation is considerably simpler than the convolution integral
that characterizes the response of linear continuous-time systems. Nevertheless,
it is a fairly complex operation, especially if the output sequence is required over
a long time period. Theorem 2.3 shows how the convolution summation can be
avoided by z-transformation.
2.6 The time response of a discrete-time system
THEOREM 2.3: THE CONVOLUTION THEOREM
The z-transform of the convolution of two time sequences is equal to the product of their
z-transforms.
PROOF
z-transforming (2.24) gives
N
X
yðkÞz2k
k50 "
#
N X
N
X
5
uðiÞhðk 2 iÞ z2k
YðzÞ 5
k50
(2.28)
i50
Interchange the order of summation and substitute j 5 k 2 i to obtain
YðzÞ 5
N X
N
X
uðiÞhðjÞz2ði1jÞ
(2.29)
i50 j52i
Using the causality property, (2.24) reduces (2.29) to
"
#"
#
N
N
X
X
2i
2j
uðiÞz
hðjÞz
YðzÞ 5
i50
(2.30)
j50
Therefore,
YðzÞ 5 HðzÞUðzÞ
(2.31)
The function H(z) of (2.31) is known as the z-transfer function or simply the
transfer function. It plays an important role in obtaining the response of an
LTI system to any input, as explained later. Note that the transfer function and
impulse response sequence are z-transform pairs.
Applying the convolution theorem to the response of an LTI system allows us to
use the z-transform to find the output of a system without convolution as follows:
1. z-transform the input.
2. Multiply the z-transform of the input and the z-transfer function.
3. Inverse z-transform to obtain the output temporal sequence.
An added advantage of this approach is that the output can often be obtained in
closed form. The preceding procedure is demonstrated in Example 2.20.
EXAMPLE 2.20
Given the discrete-time system
yðk 1 1Þ 2 0:5yðkÞ 5 uðkÞ; yð0Þ 5 0
find the impulse response of the system h(k):
1. From the difference equation
2. Using z-transformation
35
36
CHAPTER 2 Discrete-Time Systems
Solution
1. Let u(k) 5 δ(k). Then
yð1Þ
yð2Þ 5 0:5yð1Þ 5 0:5
yð3Þ 5 0:5yð2Þ 5 ð0:5Þ2
(
ð0:5Þi21 ; i 5 1; 2; 3; . . .
i:e:; hðiÞ 5
0;
i,1
2. Alternatively, z-transforming the difference equation yields the transfer function
HðzÞ 5
YðzÞ
1
5
UðzÞ z 2 0:5
Inverse-transforming with the delay theorem gives the impulse response
hðiÞ 5
ð0:5Þi21 ; i 5 1; 2; 3; . . .
0;
i,1
Observe that the response decays exponentially because the pole has magnitude less than
unity. In Chapter 4, we discuss this property and relate to the stability of discrete-time
systems.
EXAMPLE 2.21
Given the discrete-time system
yðk 1 1Þ 2 yðkÞ 5 uðk 1 1Þ
find the system transfer function and its response to a sampled unit step.
Solution
The transfer function corresponding to the difference equation is
HðzÞ 5
z
z21
We multiply the transfer function by the sampled unit step’s z-transform to obtain
YðzÞ 5
z
z
z 2
z
5z
3
5
z21
z21
z21
ðz21Þ2
The z-transform of a unit ramp is
FðzÞ 5
z
ðz21Þ2
Then, using the time advance property of the z-transform, we have the inverse transform
yðiÞ 5
i 1 1; i 5 0; 1; 2; 3; . . .
0;
i,0
2.7 The modified z-transform
It is obvious from Example 2.21 that z-transforming yields the response of a
system in closed form more easily than direct evaluation. For higher-order difference equations, obtaining the response in closed form directly may be impossible,
whereas z-transforming to obtain the response remains a relatively simple task.
2.7 The modified z-transform
Sampling and z-transformation capture the values of a continuous-time function
at the sampling points only. To evaluate the time function between sampling
points, we need to delay the sampled waveform by a fraction of a sampling interval before sampling. We can then vary the sampling points by changing the delay
period. The z-transform associated with the delayed waveform is known as the
modified z-transform.
We consider a causal continuous-time function y(t) sampled every T seconds.
Next, we insert a delay Td , T before the sampler as shown in Figure 2.7. The
output of the delay element is the waveform
yd ðtÞ 5
yðt 2 Td Þ; t $ 0
0;
t,0
(2.32)
Note that delaying a causal sequence always results in an initial zero value.
To avoid inappropriate initial values, we rewrite the delay as
Td 5 T 2 mT;
m 5 1 2 Td =T
0#m,1
(2.33)
For example, a time delay of 0.2 s with a sampling period T of 1 s corresponds
to m 5 0.8—that is, a time advance of 0.8 of a sampling period and a time delay
of one sampling period. If y21(t 1 mT) is defined as y(t 1 mT) delayed by one
complete sampling period, then, based on (2.32), yd(t) is given by
yd ðtÞ 5 yðt 2 T 1 mTÞ 5 y21 ðt 1 mTÞ
(2.34)
We now sample the delayed waveform with sampling period T to obtain
yd ðkTÞ 5 y21 ðkT 1 mTÞ;
y(t)
Delay Td
FIGURE 2.7
Sampling of a delayed signal.
yd (t)
k 5 0; 1; 2; . . .
T
yd (kT )
(2.35)
37
38
CHAPTER 2 Discrete-Time Systems
From the delay theorem, we know the z-transform of y21(t)
Y21 ðzÞ 5 z21 YðzÞ
(2.36)
We need to determine the effect of the time advance by mT to obtain the z-transform
of yd(t). We determine this effect by considering specific examples. At this point, it
suffices to write
Yðz; mÞ 5 Zm fyðkTÞg 5 z21 ZfyðkT 1 mTÞg
(2.37)
where Zm fg denotes the modified z-transform.
EXAMPLE 2.22: STEP
The step function has fixed amplitude for all time arguments. Thus, shifting it or delaying
it does not change the sampled values. We conclude that the modified z-transform of a
sampled step is the same as its z-transform, times z21 for all values of the time advance
mT—that is, 1/(1 2 z21).
EXAMPLE 2.23: EXPONENTIAL
We consider the exponential waveform
yðtÞ 5 e2pt
(2.38)
The effect of a time advance mT on the sampled values for an exponential decay is shown
in Figure 2.8. The sampled values are given by
yðkT 1 mTÞ 5 e2pðk1mÞT 5 e2pmT e2pkT ;
k 5 0; 1; 2; . . .
(2.39)
We observe that the time advance results in a scaling of the waveform by the factor e2pmT.
By the linearity of the z-transform, we have the following:
Z yðkT 1 mTÞ 5 e2pmT
z
z 2 e2pT
(2.40)
e–pkT
e–p(k+m–1)T
......
e–p(k+m)T
0
T
2T
3T
FIGURE 2.8
Effect of time advance on sampling an exponential decay.
kT
2.8 Frequency response of discrete-time systems
Using (2.37), we have the modified z-transform
e2pmT
z 2 e2pT
Yðz; mÞ 5
(2.41)
For example, for p 5 4 and T 5 0.2 s, to delay by 0.7 T, we let m 5 0.3 and calculate
e2pmT 5 e20.24 5 0.787 and e2pT 5 e20.8 5 0.449. We have the modified z-transform
Yðz; mÞ 5
0:787
z 2 0:449
The modified z-transforms of other important functions, such as the ramp and the
sinusoid, can be obtained following the procedure presented earlier. The derivations of
these modified z-transforms are left as exercises.
2.8 Frequency response of discrete-time systems
In this section, we discuss the steady-state response of a discrete-time system to
a sampled sinusoidal input.3 It is shown that, as in the continuous-time case, the
response is a sinusoid of the same frequency as the input with frequency-dependent
phase shift and magnitude scaling. The scale factor and phase shift define a complex function of frequency known as the frequency response.
We first obtain the frequency response using impulse sampling and the
Laplace transform to exploit the well-known relationship between the transfer
function Ha(s) and the frequency response Ha( jω).
Ha ðjωÞ 5 Ha ðsÞjs5jω
(2.42)
The impulse-sampled representation of a discrete-time waveform is
u ðtÞ 5
N
X
uðkTÞδðt 2 kTÞ
(2.43)
k50
where u(kT) is the value at time kT, and δ(t 2 kT) denotes a Dirac delta at time
kT. The representation (2.43) allows Laplace transformation to obtain
U ðsÞ 5
N
X
uðkTÞe2kTs
(2.44)
k50
3
Unlike continuous sinusoids, sampled sinusoids are only periodic if the ratio of the period of the
waveform and the sampling period is a rational number (equal to a ratio of integers). However, the
continuous envelope of the sampled form is clearly always periodic. See the text by Oppenheim
et al., 1997, p. 26, for more details.
39
40
CHAPTER 2 Discrete-Time Systems
It is now possible to define a transfer function for sampled inputs as
Y ðsÞ H ðsÞ 5 U ðsÞ zero initial conditions
(2.45)
Then, using (2.42), we obtain
H ðjωÞ 5 H ðsÞjs5jω
(2.46)
To rewrite (2.46) in terms of the complex variable z 5 e , we use the equation
HðzÞ 5 H ðsÞ 1
(2.47)
sT
s5T lnðzÞ
Thus, the frequency response is given by
H ðjωÞ 5 HðzÞz5ejωT
(2.48)
5 HðejωT Þ
Equation (2.48) can also be verified without the use of impulse sampling by
considering the sampled complex exponential
uðkTÞ 5 u0 ejkω0 T
5 u0 ½cosðkω0 TÞ 1 jsinðkω0 TÞ;
k 5 0; 1; 2; . . .
(2.49)
This eventually yields the sinusoidal response while avoiding its second-order
z-transform. The z-transform of the chosen input sequence is the first-order function
z
(2.50)
UðzÞ 5 u0
z 2 ejω0 T
Assume the system z-transfer function to be
HðzÞ 5
NðzÞ
n
(2.51)
L ðz 2 pi Þ
i51
where N(z) is a numerator polynomial of order n or less, and the system poles pi
are assumed to lie inside the unit circle.
The system output due to the input of (2.49) has the z-transform
2
3
6
YðzÞ 5 6
4
7
z
7 u0
5 z 2 ejω0 T
L ðz 2 pi Þ
NðzÞ
n
(2.52)
i51
This can be expanded into the partial fractions
YðzÞ 5
n
X
Az
Bi z
1
z 2 ejω0 T
z
2 pi
i51
(2.53)
2.8 Frequency response of discrete-time systems
Then inverse z-transforming gives the output
yðkTÞ 5 Aejkω0 T 1
n
X
Bi pki ;
k 5 0; 1; 2; . . .
(2.54)
i51
The assumption of poles inside the unit circle implies that, for sufficiently large k,
the output reduces to
yss ðkTÞ 5 Aejkω0 T ;
k large
(2.55)
where yss (kT) denotes the steady-state output.
The term A is the partial fraction coefficient
YðzÞ
jω0 T ðz 2 e Þ
A5
z
z5ejω0 T
(2.56)
5 Hðejω0 T Þu0
Thus, we write the steady-state output in the form
jω0 T
yss ðkTÞ 5 Hðejω0 T Þu0 ej½kω0 T1+Hðe Þ ; k large
(2.57)
The real part of this response is the response due to a sampled cosine input, and
the imaginary part is the response of a sampled sine. The sampled cosine
response is
yss ðkTÞ 5 Hðejω0 T Þu0 cos½kω0 T 1 +Hðejω0 T Þ; k large
(2.58)
The sampled sine response is similar to (2.58) with the cosine replaced by
sine.
Equations (2.57) and (2.58) show that the response to a sampled sinusoid is a
sinusoid of the same frequency scaled and phase-shifted by the magnitude and
angle
jHðejω0 T Þj+Hðejω0 T Þ
(2.59)
respectively. This is the frequency response function obtained earlier using
impulse sampling. Thus, one can use complex arithmetic to determine the steadystate response due to a sampled sinusoid without the need for z-transformation.
EXAMPLE 2.24
Find the steady-state response of the system
HðzÞ 5
1
ðz 2 0:1Þðz 2 :5Þ
due to the sampled sinusoid u(kT) 5 3 cos(0.2 k).
41
42
CHAPTER 2 Discrete-Time Systems
Solution
Using (2.58) gives the response
yss ðkTÞ 5 jHðej0:2 Þj3cosð0:2k 1 +Hðej0:2 ÞÞ; k large
1
1
3cos 0:2k 1 +
5 j0:2
ðe 2 0:1Þðej0:2 2 0:5Þ ðej0:2 2 0:1Þðej0:2 2 0:5Þ
5 6:4cosð0:2k 2 0:614Þ
2.8.1 Properties of the frequency response
of discrete-time systems
Using (2.48), the following frequency response properties can be derived:
1. DC gain: The DC gain is equal to H(1).
PROOF
From (2.48),
HðejωT Þjω-0 5 HðzÞjz-1
5 Hð1Þ
2. Periodic nature: The frequency response is a periodic function of frequency
with period ωs 5 2π/T rad/s.
PROOF
The complex exponential
ejωT 5 cosðωTÞ 1 jsinðωTÞ
is periodic with period ωs 5 2π/T rad/s. Because H(ejwT) is a single-valued function of its
argument, it follows that it also is periodic and that it has the same repetition frequency.
3. Symmetry: For transfer functions with real coefficients, the magnitude of the
transfer function is an even function of frequency and its phase is an odd
function of frequency.
PROOF
For negative frequencies, the transfer function is
Hðe2jωT Þ 5 H ejωT
For real coefficients, we have
HðejωT Þ 5 H ejωT
2.8 Frequency response of discrete-time systems
Combining the last two equations gives
Hðe2jωT Þ 5 HðejωT Þ
Equivalently, we have
Hðe2jωT Þ 5 HðejωT Þ
+Hðe2jωT Þ 5 2+HðejωT Þ
Hence, it is only necessary to obtain H(ejωT) for frequencies ω in the range
from DC to ωs/2. The frequency response for negative frequencies can be
obtained by symmetry, and for frequencies above ωs/2 the frequency response is
periodically repeated. If the frequency response has negligible amplitudes at
frequencies above ωs/2, the repeated frequency response cycles do not overlap.
The overall effect of sampling for such systems is to produce a periodic repetition
of the frequency response of a continuous-time system.
Because the frequency response functions of physical systems are not bandlimited, overlapping of the repeated frequency response cycles, known as folding,
occurs. The frequency ωs/2 is known as the folding frequency. Folding results
in distortion of the frequency response and should be minimized. This can be
accomplished by proper choice of the sampling frequency ωs/2 or filtering.
Figure 2.9 shows the magnitude of the frequency response of a second-order
underdamped digital system.
3
2.5
|H( jω)|
2
1.5
1
0.5
0
0
200
400
600
ω rad/s
800
FIGURE 2.9
Magnitude of the frequency response of a digital system.
1000
1200
43
44
CHAPTER 2 Discrete-Time Systems
2.8.2 MATLAB commands for the discrete-time
frequency response
The MATLAB commands bode, nyquist, and nichols calculate and plot the frequency response of a discrete-time system. For a sampling period of 0.2 s and a
transfer function with numerator num and denominator den, the three commands
have the form
.. g 5 tfðnum;den; 0:2Þ
.. bodeðgÞ
.. nyquistðgÞ
.. nicholsðgÞ
MATLAB limits the user’s options for automatically generated plots. However,
all the commands have alternative forms that allow the user to obtain the frequency
response data for later plotting.
The commands bode and nichols have the alternative form
.. ½M; P; w 5 bodeðg; wÞ
.. ½M; P; w 5 nicholsðg; wÞ
where w is predefined frequency grid, M is the magnitude, and P is the phase of
the frequency response. MATLAB selects the frequency grid if none is given and
returns the same list of outputs. The frequency vector can also be eliminated from
the output or replaced by a scalar for single-frequency computations. The command nyquist can take similar forms to those just described but yields the real
and imaginary parts of the frequency response as follows:
.. ½Real; Imag; w 5 nyquistðg; wÞ
As with all MATLAB commands, printing the output is suppressed if any of the
frequency response commands are followed by a semicolon. The output can
then be used with the command plot to obtain user-selected plot specifications.
For example, a plot of the actual frequency response points without connections
is obtained with the command
.. plotðRealð:Þ; Imagð:Þ; ' 'Þ
where the locations of the data points are indicated with the ' '. The command
.. subplotð2; 3; 4Þ
creates a 2-row, 3-column grid and draw axes at the first position of the second
row (the first three plots are in the first row, and 4 is the plot number). The next
plot command superimposes the plot on these axes. For other plots, the subplot
and plot commands are repeated with the appropriate arguments. For example, a
plot in the first row and second column of the grid is obtained with the command
.. subplotð2; 3; 2Þ
2.9 The sampling theorem
2.9 The sampling theorem
Sampling is necessary for the processing of analog data using digital elements.
Successful digital data processing requires that the samples reflect the nature
of the analog signal and that analog signals be recoverable, at least in theory,
from a sequence of samples. Figure 2.10 shows two distinct waveforms with identical samples. Obviously, faster sampling of the two waveforms would produce
distinguishable sequences. Thus, it is obvious that sufficiently fast sampling is
a prerequisite for successful digital data processing. The sampling theorem
gives a lower bound on the sampling rate necessary for a given band-limited
signal (i.e., a signal with a known finite bandwidth).
FIGURE 2.10
Two different waveforms with identical samples.
THEOREM 2.4: THE SAMPLING THEOREM
The band-limited signal with
F
f ðtÞ ! FðjωÞ; FðjωÞ 6¼ 0; 2ωm # ω # ωm
FðjωÞ 5 0; elsewhere
(2.60)
with F denoting the Fourier transform, can be reconstructed from the discrete-time waveform
f ðtÞ 5
N
X
f ðtÞδðt 2 kTÞ
(2.61)
k52N
if and only if the sampling angular frequency ωs 5 2π/T satisfies the condition
ωs . 2ωm
(2.62)
The spectrum of the continuous-time waveform can be recovered using an ideal low-pass
filter of bandwidth ωb in the range
ωm , ωb , ωs =2
(2.63)
PROOF
Consider the unit impulse train
δT ðtÞ 5
N
X
k52N
δðt 2 kTÞ
(2.64)
45
46
CHAPTER 2 Discrete-Time Systems
and its Fourier transform
δT ðωÞ 5
N
2π X
δðω 2 nωs Þ
T n52N
(2.65)
Impulse sampling is achieved by multiplying the waveforms f (t) and δT (t). By the frequency convolution theorem, the spectrum of the product of the two waveforms is given by
the convolution of their two spectra; that is,
F fδT ðtÞ 3 f ðtÞg 5
5
5
1
δT ðjωÞ FðjωÞ
2π
N
1 X
δðω 2 nωs Þ FðjωÞ
T n52N
N
1 X
Fðω 2 nωs Þ
T n52N
where ωm is the bandwidth of the signal. Therefore, the spectrum of the sampled waveform
is a periodic function of frequency ωs. Assuming that f(t) is a real valued function, then it
is well known that the magnitude jF(jω)j is an even function of frequency, whereas the
phase +F(jω) is an odd function. For a band-limited function, the amplitude and phase
in the frequency range 0 to ωs/2 can be recovered by an ideal low-pass filter as shown in
Figure 2.11.
2.9.1 Selection of the sampling frequency
In practice, finite bandwidth is an idealization associated with infinite-duration signals, whereas finite duration implies infinite bandwidth. To show this, assume that
a given signal is to be band limited. Band limiting is equivalent to multiplication
by a pulse in the frequency domain. By the convolution theorem, multiplication in
the frequency domain is equivalent to convolution of the inverse Fourier transforms. Hence, the inverse transform of the band-limited function is the convolution
of the original time function with the sinc function, a function of infinite duration.
We conclude that a band-limited function is of infinite duration.
–ωs
–ωs
2
ωm
2
ωb
FIGURE 2.11
Sampling theorem.
ωs
ωs
2.9 The sampling theorem
A time-limited function is the product of a function of infinite duration and
a pulse. The frequency convolution theorem states that multiplication in the time
domain is equivalent to convolution of the Fourier transforms in the frequency
domain. Thus, the spectrum of a time-limited function is the convolution of the spectrum of the function of infinite duration with a sinc function, a function of infinite
bandwidth. Hence, the Fourier transform of a time-limited function has infinite bandwidth. Because all measurements are made over a finite time period, infinite
bandwidths are unavoidable. Nevertheless, a given signal often has a finite “effective
bandwidth” beyond which its spectral components are negligible. This allows us to
treat physical signals as band limited and choose a suitable sampling rate for them
based on the sampling theorem.
In practice, the sampling rate chosen is often larger than the lower bound
specified in the sampling theorem. A rule of thumb is to choose ωs as
ωs 5 kωm ;
5 # k # 10
(2.66)
The choice of the constant k depends on the application. In many applications,
the upper bound on the sampling frequency is well below the capabilities of
state-of-the-art hardware. A closed-loop control system cannot have a sampling
period below the minimum time required for the output measurement; that is, the
sampling frequency is upper-bounded by the sensor delay.4 For example, oxygen
sensors used in automotive air/fuel ratio control have a sensor delay of about 20 ms,
which corresponds to a sampling frequency upper bound of 50 Hz. Another limitation is the computational time needed to update the control. This is becoming less
restrictive with the availability of faster microprocessors but must be considered
in sampling rate selection.
In digital control, the sampling frequency must be chosen so that samples provide
a good representation of the analog physical variables. A more detailed discussion of
the practical issues that must be considered when choosing the sampling frequency is
given in Chapter 12. Here, we only discuss choosing the sampling period based on
the sampling theorem.
For a linear system, the output of the system has a spectrum given by the product of the frequency response and input spectrum. Because the input is not known a
priori, we must base our choice of sampling frequency on the frequency response.
The frequency response of a first-order system is
HðjωÞ 5
K
jω=ωb 1 1
(2.67)
where K is the DC gain and ωb is the system bandwidth. The frequency response
amplitude drops below the DC level by a factor of about 10 at the frequency 7ωb.
If we consider ωm 5 7ωb, the sampling frequency is chosen as
ωs 5 kωb ; 35 # k # 70
4
(2.68)
It is possible to have the sensor delay as an integer multiple of the sampling period if a state estimator is used, as discussed in Franklin et al. (1998).
47
48
CHAPTER 2 Discrete-Time Systems
For a second-order system with frequency response
HðjωÞ 5
K
j2ζω=ωn 1 1 2 ðω=ωn Þ2
(2.69)
the bandwidth of the system is approximated by the damped natural frequency
qffiffiffiffiffiffiffiffiffiffiffiffiffi
(2.70)
ωd 5 ωn 1 2 ζ 2
Using a frequency of 7ωd as the maximum significant frequency, we choose
the sampling frequency as
ωs 5 kωd ; 35 # k # 70
(2.71)
In addition, the impulse response of a second-order system is of the form
yðtÞ 5 Ae2ζωn t sinðωd t 1 φÞ
(2.72)
where A is a constant amplitude and φ is a phase angle. Thus, the choice of
sampling frequency of (2.71) is sufficiently fast for oscillations of frequency ωd
and time to first peak π/ωd.
EXAMPLE 2.25
Given a first-order system of bandwidth 10 rad/s, select a suitable sampling frequency and
find the corresponding sampling period.
Solution
A suitable choice of sampling frequency is ωs 5 60, ωd 5 600 rad/s. The corresponding
sampling period is approximately T 5 2π/ωs 0.01 s.
EXAMPLE 2.26
A closed-loop control system must be designed for a steady-state error not to exceed 5 percent, a damping ratio of about 0.7, and an undamped natural frequency of 10 rad/s.
Select a suitable sampling period for the system if the system has a sensor delay of
1. 0.02 s
2. 0.03 s
Solution
Let the sampling frequency be
ωs $ 35ωd
pffiffiffiffiffiffiffiffiffiffiffiffiffi
5 35ωn 1 2 ζ 2
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
5 350 1 2 0:49
5 249:95 rad=s
The corresponding sampling period is T 5 2π/ωs # 0.025 s.
1. A suitable choice is T 5 20 ms because this is equal to the sensor delay.
2. We are forced to choose T 5 30 ms, which is equal to the sensor delay.
Problems
Resources
Chen, C.-T., 1989. System and Signal Analysis. Saunders, New York.
Feuer, A., Goodwin, G.C., 1996. Sampling in Digital Signal Processing and Control.
Birkhauser, Boston.
Franklin, G.F., Powell, J.D., Workman, M.L., 1998. Digital Control of Dynamic Systems.
Addison-Wesley, Menlo Park, CA.
Goldberg, S., 1986. Introduction to Difference Equations. Dover, Mineola, NY.
Jacquot, R.G., 1981. Modern Digital Control Systems. Marcel Dekker, New York.
Kuo, B.C., 1992. Digital Control Systems. Saunders, Ft. Worth, TX.
Mickens, R.E., 1987. Difference Equations. Van Nostrand Reinhold, New York.
Oppenheim, A.V., Willsky, A.S., Nawab, S.H., 1997. Signals and Systems. Prentice Hall,
Englewood Cliffs, NJ.
PROBLEMS
2.1
Derive the discrete-time model of Example 2.1 from the solution of the
system differential equation with initial time kT and final time (k 1 1)T.
2.2
For each of the following equations, determine the order of the equation
and then test it for (i) linearity, (ii) time invariance, and (iii) homogeneity.
(a) y(k 1 2) 5 y(k 1 1) y(k) 1 u(k)
(b) y(k 1 3) 1 2 y(k) 5 0
(c) y(k 1 4) 1 y(k 2 1) 5 u(k)
(d) y(k 1 5) 5 y(k 1 4) 1 u(k 1 1) 2 u(k)
(e) y(k 1 2) 5 y(k) u(k)
2.3
Find the transforms of the following sequences using Definition 2.1.
(a) {0, 1, 2, 4, 0, 0, . . .}
(b) {0, 0, 0, 1, 1, 1, 0, 0, 0, . . .}
(c) {0, 220.5, 1, 220.5, 0, 0, 0, . . .}
2.4
Obtain closed forms of the transforms of Problem 2.3 using the table of
z-transforms and the time delay property.
2.5
Prove the linearity and time delay properties of the z-transform from basic
principles.
2.6
Use the linearity of the z-transform and the transform of the exponential
function to obtain the transforms of the discrete-time functions.
(a) sin(kωT)
(b) cos(kωT)
2.7
Use the multiplication by exponential property to obtain the transforms of
the discrete-time functions.
(a) e2αkTsin(kωT)
(b) e2αkTcos(kωT)
49
50
CHAPTER 2 Discrete-Time Systems
2.8
Find the inverse transforms of the following functions using Definition 2.1
and, if necessary, long division.
(a) F(z) 5 1 1 3z21 1 4z22
(b) F(z) 5 5z21 1 4z25
z
(c)
FðzÞ 5 2
z 1 0:3z 1 0:02
(d)
2.9
FðzÞ 5
z2
z 2 0:1
1 0:04z 1 0:25
For Problems 2.8(c) and (d), find the inverse transforms of the functions
using partial fraction expansion and table lookup.
2.10
Solve the following difference equations.
(a) y(k 1 1) 2 0.8 y(k) 5 0, y(0) 5 1
(b) y(k 1 1) 2 0.8 y(k) 5 1(k), y(0) 5 0
(c) y(k 1 1) 2 0.8 y(k) 5 1(k), y(0) 5 1
(d) y(k 1 2) 1 0.7 y(k 1 1) 1 0.06 y(k) 5 δ(k), y(0) 5 0, y(1) 5 2
2.11
Find the transfer functions corresponding to the difference equations of
Problem 2.2 with input u(k) and output y(k). If no transfer function is
defined, explain why.
2.12
Test the linearity with respect to the input of the systems for which you
found transfer functions in Problem 2.11.
2.13
If the rational functions of Problems 2.8(c) and (d) are transfer functions
of LTI systems, find the difference equation governing each system.
2.14
We can use z-transforms to find the sum of integers raised to various
powers. This is accomplished by first recognizing that the sum is the solution of the difference equation
f ðkÞ 5 f ðk 2 1Þ 1 aðkÞ
where a(k) is the kth term in the summation. Evaluate the following summations using z-transforms.
n
P
k
(a)
k51
(b)
n
P
k2
k51
2.15
Given the discrete-time system
yðk 1 2Þ 2 yðkÞ 5 2uðkÞ
Problems
find the impulse response of the system gðkÞ:
(a) From the difference equation
(b) Using z-transformation
2.16
The following identity provides a recursion for the cosine function:
cosðkxÞ 5 2cosððk 2 1ÞxÞcosððk 2 2ÞxÞ;
k integer
To verify its validity, let f ðkÞ 5 cosðkxÞ and rewrite the expression as a
difference equation. Show that the solution of the difference equation is
indeed
f ðkÞ 5 cosðkxÞ
2.17
Repeat Problem 2.16 for the identity
sinðkxÞ 5 2sinððk 2 1ÞxÞcosðxÞ 2 sinððk 2 2ÞxÞ;
k integer
2.18
Find the impulse response functions for the systems governed by the following difference equations.
(a) y(k 1 1) 2 0.5 y(k) 5 u(k)
(b) y(k 1 2) 2 0.1 y(k 1 1) 1 0.8 y(k) 5 u(k)
2.19
Find the final value for the functions if it exists.
z
(a) FðzÞ 5 2
z 2 1:2z 1 0:2
z
(b) FðzÞ 5 2
z 1 0:3z 1 2
2.20
Find the steady-state response of the systems resulting from the sinusoidal
input u(k) 5 0.5 sin(0.4 k).
z
(a) HðzÞ 5
z 2 0:4
z
(b) HðzÞ 5 2
z 1 0:4z 1 0:03
2.21
Find the frequency response of a noncausal system whose impulse
response sequence is given by
fuðkÞ; uðkÞ 5 uðk 1 KÞ; k 5 2N; . . .; Ng
Hint: The impulse response sequence is periodic with period K and can be
expressed as
K 21 X
N
X
u ðtÞ 5
uðl 1 mKÞδðt 2 l 2 mKÞ
l50 m52N
51
52
CHAPTER 2 Discrete-Time Systems
2.22
The well-known Shannon reconstruction theorem states that any band-limited signal u(t) with bandwidth ωs/2 can be exactly reconstructed from its
samples at a rate ωs 5 2π/T. The reconstruction is given by
hω
i
s
ðt 2 kTÞ
sin
uðkÞ ωs2
uðtÞ 5
ðt 2 kTÞ
k52N
2
N
X
Use the convolution theorem to justify the preceding expression.
2.23
Obtain the convolution of the two sequences {1, 1, 1} and {1, 2, 3}.
(a) Directly
(b) Using z-transformation
2.24
Obtain the modified z-transforms for the functions of Problems (2.6) and (2.7).
2.25
Using the modified z-transform, examine the intersample behavior of the functions h(k) of Problem 2.18. Use delays of (1) 0.3T, (2) 0.5T, and (3) 0.8T.
Attempt to obtain the modified z-transform for Problem 2.19 and explain
why it is not defined.
2.26
The following open-loop systems are to be digitally feedback-controlled.
Select a suitable sampling period for each if the closed-loop system is to
be designed for the given specifications.
1
Time constant 5 0.1 s
(a) Gol ðsÞ 5
s13
1
Undamped natural frequency 5 5 rad/s, damping
(b) Gol ðsÞ 5 2
s 1 4s 1 3
ratio 5 0.7
2.27
Repeat Problem 2.26 if the systems have the following sensor delays.
(a) 0.025 s
(b) 0.03 s
COMPUTER EXERCISES
2.28
Consider the closed-loop system of Problem 2.26(a).
(a) Find the impulse response of the closed-loop transfer function, and
obtain the impulse response sequence for a sampled system output.
(b) Obtain the z-transfer function by z-transforming the impulse response
sequence.
(c) Using MATLAB, obtain the frequency response plots for sampling
frequencies ωs 5 kωb, k 5 5, 35, 70.
(d) Comment on the choices of sampling periods of part (c).
Computer Exercises
2.29
Repeat Problem 2.28 for the second-order closed-loop system of Problem
2.26(b) with plots for sampling frequencies ωs 5 kωd, k 5 5, 35, 70.
2.30
Use MATLAB with a sampling period of 1 s and a delay of 0.5 s to verify
the results of Problem 2.20 for ω 5 5 rad/s and α 5 2 s21.
2.31
The following difference equation describes the evolution of the expected
price of a commodity5
pe ðk 1 1Þ 5 ð1 2 γÞpe ðkÞ 1 γpðkÞ
where pe(k) is the expected price after k quarters, p(k) is the actual price
after k quarters, and γ is a constant.
(a) Simulate the system with γ 5 0.5 and a fixed actual price of one unit,
and plot the actual and expected prices. Discuss the accuracy of the
model prediction.
(b) Repeat part (a) for an exponentially decaying price p(k) 5 (0.4)k.
(c) Discuss the predictions of the model referring to your simulation
results.
5
Gujarate, D.N., 1988. Basic Econometrics. McGraw-Hill, NY, p. 547.
53
CHAPTER
Modeling of Digital Control
Systems
3
OBJECTIVES
After completing this chapter, the reader will be able to do the following:
1. Obtain the transfer function of an analog system with analog-to-digital and digitalto-analog converters, including systems with a time delay.
2. Find the closed-loop transfer function for a digital control system.
3. Find the steady-state tracking error for a closed-loop control system.
4. Find the steady-state error caused by a disturbance input for a closed-loop control
system.
As in the case of analog control, mathematical models are needed for the analysis
and design of digital control systems. A common configuration for digital control
systems is shown in Figure 3.1. The configuration includes a digital-to-analog
converter (DAC), an analog subsystem, and an analog-to-digital converter (ADC).
The DAC converts numbers calculated by a microprocessor or computer into
analog electrical signals that can be amplified and used to control an analog plant.
The analog subsystem includes the plant as well as the amplifiers and actuators
necessary to drive it. The output of the plant is periodically measured and converted to a number that can be fed back to the computer using an ADC. In this
chapter, we develop models for the various components of this digital control
configuration. Many other configurations that include the same components can
be similarly analyzed. We begin by developing models for the ADC and DAC,
then for the combination of DAC, analog subsystem, and ADC.
3.1 ADC model
Assume that
•
•
•
ADC outputs are exactly equal in magnitude to their inputs (i.e., quantization
errors are negligible).
The ADC yields a digital output instantaneously.
Sampling is perfectly uniform (i.e., occurs at a fixed rate).
Then the ADC can be modeled as an ideal sampler with sampling period T as
shown in Figure 3.2.
55
56
CHAPTER 3 Modeling of Digital Control Systems
External
Input
Computer or
Microprocessor
Analog
Subsystem
DAC
Analog
Output
ADC
FIGURE 3.1
Common digital control system configuration.
T
FIGURE 3.2
Ideal sampler model of an ADC.
Clearly, the preceding assumptions are idealizations that can only be approximately true in practice. Quantization errors are typically small but nonzero; variations in sampling rate occur but are negligible, and physical ADCs have a finite
conversion time. Nevertheless, the ideal sampler model is acceptable for most
engineering applications.
3.2 DAC model
Assume that
•
•
•
DAC outputs are exactly equal in magnitude to their inputs.
The DAC yields an analog output instantaneously.
DAC outputs are constant over each sampling period.
Then the input-output relationship of the DAC is given by
ZOH
fuðkÞg ! uðtÞ 5 uðkÞ; kT # t , ðk 1 1ÞT;
k 5 0; 1; 2; . . .
(3.1)
where {u(k)} is the input sequence. This equation describes a zero-order hold
(ZOH), shown in Figure 3.3. Other functions may also be used to construct an
analog signal from a sequence of numbers. For example, a first-order hold constructs analog signals in terms of straight lines, whereas a second-order hold
constructs them in terms of parabolas.
In practice, the DAC requires a short but nonzero interval to yield an output;
its output is not exactly equal in magnitude to its input and may vary slightly
3.3 The transfer function of the ZOH
kT
kT
Zero-Order
Hold
(k+1)T
Positive Step
Negative Step
FIGURE 3.3
Model of a DAC as a zero-order hold.
over a sampling period. But the model of (3.1) is sufficiently accurate for most
engineering applications. The zero-order hold is the most commonly used DAC
model and is adopted in most digital control texts. Analyses involving other hold
circuits are similar, as seen from Problem 3.2.
3.3 The transfer function of the ZOH
To obtain the transfer function of the ZOH, we replace the number or discrete
impulse shown in Figure 3.3 by an impulse δ(t). The transfer function can then
be obtained by Laplace transformation of the impulse response. As shown in the
figure, the impulse response is a unit pulse of width T. A pulse can be represented
as a positive step at time zero followed by a negative step at time T. Using the
Laplace transform of a unit step and the time delay theorem for Laplace
transforms,
1
s
e2sT
L f1ðt 2 TÞg 5
s
L f1ðtÞg 5
(3.2)
where 1(t) denotes a unit step.
Thus, the transfer function of the ZOH is
GZOH ðsÞ 5
12 e2sT
s
(3.3)
Next, we consider the frequency response of the ZOH:
GZOH ðjωÞ 5
12 e2jωT
jω
(3.4)
57
58
CHAPTER 3 Modeling of Digital Control Systems
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
–30
–20
–10
0
10
20
30
FIGURE 3.4
Magnitude of the frequency response of the zero-order hold with T 5 1 s.
We rewrite the frequency response in the form
!
2jωT
jωT
2jωT
e 2 e 2 2e 2
GZOH ðjωÞ 5
ω
j
T
sin
ω
2jωT 2
2jωT
e 2
T
2sin ω
5
5 Te 2
T
ω
2
ω
2
We now have
G ZOH ðjωÞ+GZOH ðjωÞ 5 T sinc ωT + 2 ωT ; 2 2π , ω , 2π
2 2
T
T
(3.5)
In the frequency range of interest where the sinc function is positive, the angle of
frequency response of the ZOH hold is seen to decrease linearly with frequency,
whereas the magnitude is proportional to the sinc function. As shown in Figure 3.4,
the magnitude is oscillatory with its peak magnitude equal to the sampling period
and occurring at the zero frequency.
3.4 Effect of the sampler on the transfer function
of a cascade
In a discrete-time system including several analog subsystems in cascade and several samplers, the location of the sampler plays an important role in determining
3.4 Effect of the sampler on the transfer function of a cascade
the overall transfer function. Assuming that interconnection does not change the
mathematical models of the subsystems, the Laplace transform of the output of
the system of Figure 3.5 is given by
Y ðsÞ 5 H2 ðsÞX ðsÞ
5 H2 ðsÞH1 ðsÞUðsÞ
(3.6)
Inverse Laplace transforming gives the time response
Ðt
yðtÞ 5 0 h2 ðt 2 τÞxðτÞdτ
Ðτ
Ðt
5 0 h2 ðt 2 τÞ 0 h1 ðτ 2 λÞuðλÞdλ dτ
(3.7)
Changing the order and variables of integration, we obtain
Ðt
Ðτ
yðtÞ 5 0 uðt 2 τÞ 0 h1 ðτ 2 λÞh2 ðλÞdλ dτ
(3.8)
Ðt
5 0 uðt 2 τÞheq ðτÞdτ
Ðt
where heq ðtÞ 5 0 h1 ðt 2 τÞh2 ðλÞdτ.
Thus, the equivalent impulse response for the cascade is given by the convolution of the cascaded impulse responses. The same conclusion can be reached by
inverse-transforming the product of the s-domain transfer functions. The time
domain expression shows more clearly that cascading results in a new form for
the impulse response. So if the output of the system is sampled to obtain
ð iT
uðiT 2 τÞheq ðτÞdτ;
i 5 1; 2; . . .
(3.9)
yðiTÞ 5
o
it is not possible to separate the three time functions that are convolved to produce it.
By contrast, convolving an impulse-sampled function u (t) with a continuoustime signal as shown in Figure 3.6 results in repetitions of the continuous-time
function, each of which is displaced to the location of an impulse in the train.
Unlike the earlier situation, the resultant time function is not entirely new,
U(s)
H1(s)
X(s)
Y(s)
H2(s)
FIGURE 3.5
Cascade of two analog systems.
U(s)
U * (s)
H(s)
T
FIGURE 3.6
Analog system with sampled input.
Y(s)
59
60
CHAPTER 3 Modeling of Digital Control Systems
and there is hope of separating the functions that produced it. For a linear timeinvariant (LTI) system with impulse-sampled input, the output is given by
Ðt
yðtÞ 5 0 hðt 2 τÞu
" ðτÞdτ
#
N
X
Ðt
(3.10)
uðkTÞδðτ 2 kTÞ dτ
5 0 hðt 2 τÞ
k50
Changing the order of summation and integration gives
ðt
N
X
uðkTÞ hðt 2 τÞδðτ 2 kTÞdτ
yðtÞ 5
0
k50
(3.11)
N
X
uðkTÞhðt 2 kTÞ
5
k50
Sampling the output yields the convolution summation
yðiTÞ 5
N
X
uðkTÞhðiT 2 kTÞ;
i 5 0; 1; 2; 3; . . .
(3.12)
k50
As discussed earlier, the convolution summation has the z-transform
YðzÞ 5 HðzÞUðzÞ
(3.13)
Y ðsÞ 5 H ðsÞU ðsÞ
(3.14)
or in s-domain notation
If a single block is an equivalent transfer function for the cascade of
Figure 3.5, then its components cannot be separated after sampling. However,
if the cascade is separated by samplers, then each block has a sampled output
and input as well as a z-domain transfer function. For n blocks not separated by
samplers, we use the notation
Y ðzÞ 5 H ðzÞUðzÞ
5 ðH1 H2 . . . Hn ÞðzÞUðzÞ
(3.15)
as opposed to n blocks separated by samplers where
Y ðzÞ 5 HðzÞUðzÞ
5 H1 ðzÞH2 ðzÞ . . . Hn ðzÞUðzÞ
(3.16)
EXAMPLE 3.1
Find the equivalent sampled impulse response sequence and the equivalent z-transfer
function for the cascade of the two analog systems with sampled input
1
2
H2 ðsÞ 5
s12
s14
1. If the systems are directly connected
2. If the systems are separated by a sampler
H1 ðsÞ 5
3.5 DAC, analog subsystem, and ADC combination transfer function
Solution
1. In the absence of samplers between the systems, the overall transfer function is
HðsÞ 5
2
ðs 1 2Þðs 1 4Þ
5
1
1
2
s12 s14
The impulse response of the cascade is
hðtÞ 5 e22t 2 e24t
and the sampled impulse response is
hðkTÞ 5 e22kT 2 e24kT ;
k 5 0; 1; 2; . . .
Thus, the z-domain transfer function is
HðzÞ 5
z
z
ðe22T 2 e24T Þz
2
5
z 2 e22T
z 2 e24T
ðz 2 e22T Þðz 2 e24T Þ
2. If the analog systems are separated by a sampler, then each has a z-domain transfer
function, and the transfer functions are given by
H1 ðzÞ 5
z
z 2 e22T
H2 ðzÞ 5
2z
z 2 e24T
The overall transfer function for the cascade is
HðzÞ 5
2z2
ðz 2 e22T Þðz 2 e24T Þ
The partial fraction expansion of the transfer function is
HðzÞ 5
2
e22T z
e24T z
2
e22T 2 e24T z 2 e22T
z 2 e24T
Inverse z-transforming gives the impulse response sequence
hðkTÞ 5
2
e22T e22kT 2 e24T e24kT
e22T 2 e24T
5
2
e22ðk11ÞT 2 e24ðk11ÞT ;
e22T 2 e24T
k 5 0; 1; 2; . . .
Example 3.1 clearly shows the effect of placing a sampler between analog
blocks on the impulse responses and the corresponding z-domain transfer function.
3.5 DAC, analog subsystem, and ADC combination
transfer function
The cascade of a DAC, analog subsystem, and ADC, shown in Figure 3.7,
appears frequently in digital control systems (see Figure 3.1, for example).
61
62
CHAPTER 3 Modeling of Digital Control Systems
T
kT
kT
Zero-Order
Hold
(k+1)T
Analog
Subsystem
Positive step at kT
kT
(k+1)T
Negative step at (k+1)T
FIGURE 3.7
Cascade of a DAC, analog subsystem, and ADC.
Because both the input and the output of the cascade are sampled, it is possible
to obtain its z-domain transfer function in terms of the transfer functions of the
individual subsystems. The transfer function is derived using the discussion of
cascades given in Section 3.4.
Using the DAC model of Section 3.3, and assuming that the transfer function
of the analog subsystem is G(s), the transfer function of the DAC and analog
subsystem cascade is
GZA ðsÞ 5 GðsÞGZOH ðsÞ
GðsÞ
5 ð1 2 e2sT Þ
s
(3.17)
The corresponding impulse response is
gZA ðtÞ 5 gðtÞ gZOH ðtÞ
5 gs ðtÞ 2 gs ðt 2 TÞ
21 G ðsÞ
gs ðtÞ 5 L
s
(3.18)
The impulse response of (3.18) is the analog system step response minus a
second step response delayed by one sampling period. This response is shown
in Figure 3.8 for a second-order underdamped analog subsystem. The analog
response of (3.18) is sampled to give the sampled impulse response
gZA ðkTÞ 5 gs ðkTÞ 2 gs ðkT 2 TÞ
(3.19)
By z-transforming, we obtain the z-transfer function of the DAC (zero-order
hold), analog subsystem, and ADC (ideal sampler) cascade
GZAS ðzÞ 5 ð1 2 z21 Þ Z fgs ðtÞg
GðsÞ 5 ð1 2 z21 Þ Z L 21
s
(3.20)
3.5 DAC, analog subsystem, and ADC combination transfer function
1.5
1.0
1.0
0.8
Positive Step
0.5
0.6
0.0
0.4
−0.5
0.2
Delayed Negative
Step
−1.0
−1.5
0.0
0
2
4
6
8
10
Time s
−0.2
0
(a)
2
4
6
8
10
Time s
(b)
FIGURE 3.8
Impulse response of a DAC and analog subsystem. (a) Response of an analog system to
step inputs. (b) Response of an analog system to a unit pulse input.
The cumbersome notation in (3.20) is used to emphasize that sampling of a
time function is necessary before z-transformation. Having made this point, the
equation can be rewritten more concisely as
GðsÞ
21
(3.21)
GZAS ðzÞ 5 ð1 2 z Þ Z
s
EXAMPLE 3.2
Find GZAS(z) for the cruise control system for the vehicle shown in Figure 3.9, where u is
the input force, v is the velocity of the car, and b is the viscous friction coefficient.
Solution
We first draw a schematic to represent the cruise control system as shown in Figure 3.10.
Using Newton’s law, we obtain the following model:
_ 1 bvðtÞ 5 uðtÞ
M vðtÞ
which corresponds to the following transfer function:
GðsÞ 5
VðsÞ
1
5
UðsÞ Ms 1 b
We rewrite the transfer function in the form
GðsÞ 5
K
K=τ
5
τs 1 1 s 1 1=τ
63
64
CHAPTER 3 Modeling of Digital Control Systems
FIGURE 3.9
Automotive vehicle.
bv
u
M
FIGURE 3.10
Schematic representation of a cruise control system for an automotive vehicle.
where K 5 1/b and τ 5 M/b. The corresponding partial fraction expansion is
GðsÞ
K
A1
A2
1
5
s
s 1 1=τ
s
τ
where
A1 5
1 5τ
s11=τ s50
1
A2 5 s
5 2τ
s521=τ
Using (3.21) and the z-transform table (see Appendix I), the desired z-domain transfer
function is
80 12
39
< K
τ
τ 5=
21
@
A
4
GZAS ðzÞ 5 ð12 z Þ Z
2
: τ
s
s 1 1=τ ;
2
3
z21 5
5 K 41 1
z 2 e2T=τ
We simplify to obtain the transfer function
GZAS ðzÞ 5 K
2z 2ð1 1 e2T=τ Þ
z 2 e2T=τ
EXAMPLE 3.3
Find GZAS(z) for the vehicle position control system shown in Figure 3.10, where u is the
input force, y is the position of the car, and b is the viscous friction coefficient.
3.5 DAC, analog subsystem, and ADC combination transfer function
Solution
As with the previous example, we obtain the following equation of motion:
_ 5 uðtÞ
My€ðtÞ 1 byðtÞ
and the corresponding transfer function
GðsÞ 5
YðsÞ
1
5
UðsÞ sðMs 1 bÞ
We rewrite the transfer function in terms of the system time constant
GðsÞ 5
K
K=τ
5
sðτs 1 1Þ sðs 1 1=τÞ
where K 5 1/b and τ 5 M/b. The corresponding partial fraction expansion is
GðsÞ
K
A11
A12
A2
1
1
5
s2
s
s 1 1=τ
s
τ
where
A11 5
1 5τ
s11=τ S50
d
1
5 2τ 2
ds s11=τ S50
1
A2 5 2 5 τ2
s s521=τ
A12 5
Using (3.21), the desired z-domain transfer function is
8 2
39
=
< A
A
A
11
12
2
5
GZAS ðzÞ 5 ð1 2 z Þ Z K 4 2 1
1
:
s
s
s 1 1=τ ;
21
2
3
T
τðz 2 1Þ 5
4
5K
2τ 1
z21
z 2 e2T=τ
which can be simplified to
"
GZAS ðzÞ 5 K
T 2 τ 1 τe2T=τ z 1 τ 2 e2T=τ ðτ 1 TÞ
ðz 2 1Þ z 2 e2T=τ
#
EXAMPLE 3.4
Find GZAS(z) for the series R-L circuit shown in Figure 3.11 with the inductor voltage as
output.
Solution
Using the voltage divider rule gives
Vo
Ls
ðL=RÞs
τs
5
5
5
Vin
R 1 Ls 1 1ðL=RÞs 1 1 τs
τ5
L
R
65
66
CHAPTER 3 Modeling of Digital Control Systems
R
Vin
L
Vo
FIGURE 3.11
Series R-L circuit.
Hence, using (3.21), we obtain
9
1 =
GZAS ðzÞ 5 ð1 2 z Þ Z
:s 1 1=τ ;
21
5
8
<
z21
z
z21
5
3
z
z 2 e2T=τ
z 2 e2T=τ
EXAMPLE 3.5
Find the z-domain transfer function of the furnace sketched in Figure 3.12, where the
inside temperature Ti is the controlled variable, Tw is the wall temperature, and To is the
outside temperature. Assume perfect insulation so that there is no heat transfer between
the wall and the environment with heating provided through a resistor. The control variable
u has the dimension of temperature scaled by an amplifier with gain K. The sampling
period T 5 1 s.
Solution
The system can be modeled with the following differential equations:
T_ w ðtÞ 5 grw ðKuðtÞ 2 Tw ðtÞÞ 1 giw ðTi ðtÞ 2 Tw ðtÞÞ
T_i ðtÞ 5 giw ðTw ðtÞ 2 Ti ðtÞÞ
To
Tw
u
FIGURE 3.12
Schematic of a furnace.
Ti
3.5 DAC, analog subsystem, and ADC combination transfer function
where grw and giw are the heat transfer coefficients. Laplace transforming and simplifying,
we obtain the following transfer function
GðsÞ 5
Ti ðsÞ
Kgrw giw
5 2
s 1ð2giw 1 grw Þs 1 grw giw
UðsÞ
Note that the two poles are real and the transfer function can be rewritten in terms of
the poles p1 and p2 as
GðsÞ 5
Y ðsÞ
K
5
U ðsÞ ðs 1 p1 Þðs 1 p2 Þ
(3.22)
The corresponding partial fraction expansion is
GðsÞ
A1
A2
A3
1
1
5K
s
s 1 p1
s 1 p2
s
where
1
5 1
A1 5
ðs1p1 Þðs1p2 Þ s50 p1 p2
1
1
A2 5
52
sðs1p2 Þ s52p1
p1 ðp2 2 p1 Þ
1
1
5
A3 5
sðs1p1 Þ s52p2 p2 ðp2 2 p1 Þ
Using (3.21), the desired z-domain transfer function is
8 2
39
=
<
1
1
1
21
4
5
2
1
GZAS ðzÞ 5 ð1 2 z Þ Z K
: p1 p2 s p1 ðp2 2 p1 Þðs 1 p1 Þ p2 ðp2 2 p1 Þðs 1 p2 Þ ;
2
3
1
z
2
1
z
2
1
5
5 K4
1
2
p1 p2
p1 ðp2 2 p1 Þðz 2 ep1 T Þ p2 ðp2 2 p1 Þðz 2 ep2 T Þ
which can be simplified to
3
2
p e2p1 T 2p2 e2p2 T 1 p2 e2ðp1 1p2 ÞT
ðp1 e2p2 T 2 p2 e2p1 T 1 p2 2 p1 Þz 1 1
7
6
2p1 e2ðp1 1p2 ÞT
7
GZAS ðzÞ 5 K 6
5
4
p1 p2 ðp2 2 p1 Þðz 2 e2p1 T Þðz 2 e2p2 T Þ
EXAMPLE 3.6
Find the z-domain transfer function of an armature-controlled DC motor.
Solution
The system can be modeled by means of the following differential equations:
_ 5 Kt iðtÞ
Jθ€ ðtÞ 1 bθðtÞ
L
diðtÞ
_
1 RiðtÞ 5 uðtÞ 2 Ke θðtÞ
dt
yðtÞ 5 θðtÞ
67
68
CHAPTER 3 Modeling of Digital Control Systems
where θ is the position of the shaft (i.e., the output y of the system), i is the armature
current, u is the source voltage (i.e., the input of the system), J is the moment of inertia
of the motor, b is the viscous friction coefficient, Kt is the torque constant, Ke is the
back e.m.f. constant, R is the electric resistance, and L is the electric inductance. Laplace
transforming and simplifying gives the following transfer function:
GðsÞ 5
YðsÞ
Kt
5
UðsÞ s½ðJs 1 bÞðLs 1 RÞ 1 Kt Ke which can be rewritten as
GðsÞ 5
YðsÞ
K
5
UðsÞ sðs 1 p1 Þðs 1 p2 Þ
with appropriate values of K, p1, and p2. The corresponding partial fraction expansion is
GðsÞ
A11
A12
A2
A3
1
1
5K 2 1
s
s
s 1 p1
s 1 p2
s
where
1
5 1
ðs1p1 Þðs1p2 Þ s50 p1 p2
2
3
d4
1
5 5 2 p1 1 p2
A12 5
ds ðs1p1 Þðs1p2 Þ p21 p22
s50
1
1
A2 5 2
5
s ðs1p2 Þ s52p1 p21 ðp2 2 p1 Þ
1
1
52 2
A3 5 2
s ðs1p1 Þ s52p2
p2 ðp2 2 p1 Þ
A11 5
Using (3.21), the desired z-domain transfer function is
8 2
39
=
<
1
p1 1p2
1
1
21
4
5
1
2
GZAS ðzÞ 5ð12z Þ Z K
2
2
2
2
2
: p1 p2 s2
p1 p2 s
p1 ðp2 2 p1 Þðs1p1 Þ p2 ðp2 2 p1 Þðs1 p2 Þ ;
2
3
T
p
1p
1
1
1
2
5
5K 4
2 2 2 1
2
p1 p2 ðz2 1Þ
p1 ðp2 2p1 Þðz2 1Þðz2 ep1 T Þ p2 ðp2 2p1 Þðz2 1Þðz 2ep2 T Þ
p1 p2
_
Note that if the velocity of the motor is considered as output (i.e., yðtÞ 5 θðtÞ),
we have
the transfer function
GðsÞ 5
YðsÞ
Kt
5
UðsÞ ðJs 1 bÞðLs 1 RÞ 1 Kt Ke
and the calculations of Example 3.5 can be repeated to obtain the z-domain transfer
function (see (3.22)).
In the preceding examples, we observe that if the analog system has a pole at pS,
then GZAS(z) has a pole at pz 5 eps T . The division by s in (3.21) results in a pole at
z 5 1 that cancels, leaving the same poles as those obtained when sampling and
z-transforming the impulse response of the analog subsystem. However, the zeros
of the transfer function are different in the presence of a DAC.
3.6 Systems with transport lag
3.6 Systems with transport lag
Many physical system models include a transport lag or delay in their transfer
functions. These include chemical processes, automotive engines, sensors, digital
systems, and so on. In this section, we obtain the z-transfer function GZAS(z) of (3.21)
for a system with transport delay.
The transfer function for systems with a transport delay is of the form
GðsÞ 5 Ga ðsÞe2Td s
(3.23)
where Td is the transport delay. As in Section 2.7, the transport delay can be
rewritten as
Td 5 lT 2 mT;
0#m,1
(3.24)
where m is a positive integer. For example, a time delay of 3.1 s with a sampling
period T of 1 s corresponds to l 5 4 and m 5 0.9. A delay by an integer multiple
of the sampling period does not affect the form of the impulse response of the
system. Therefore, using the delay theorem and (3.20), the z-transfer function for
the system of (3.23) can be rewritten as
3 9
=
2Tðl2mÞs
G
ðsÞe
a
5
GZAS ðzÞ 5 ð1 2 z21 Þ Z L 21 4
;
:
s
8
<
2
8
<
2
5 z2l ð1 2 z21 Þ Z L 21 4
:
Ga ðsÞe
s
mTs
3 9
=
5
;
(3.25)
From (3.25), we observe that the inverse Laplace transform of the function
Gs ðsÞ 5
Ga ðsÞ
s
(3.26)
is sampled and z-transformed to obtain the desired z-transfer function. We rewrite
(3.26) in terms of Gs(s) as
GZAS ðzÞ 5 z2l ð1 2 z21 Þ Zf L 21 ½Gs ðsÞemTs g
(3.27)
Using the time advance theorem of Laplace transforms gives
GZAS ðzÞ 5 z2l ð1 2 z21 Þ Zfgs ðt 1 mTÞg
(3.28)
where the impulse-sampled waveform must be used to allow Laplace transformation. The remainder of the derivation does not require impulse sampling,
69
70
CHAPTER 3 Modeling of Digital Control Systems
and we can replace the impulse-sampled waveform with the corresponding
sequence
GZAS ðzÞ 5 z2l ð1 2 z21 Þ Z fgs ðkT 1 mTÞg
(3.29)
The preceding result hinges on our ability to obtain the effect of a time advance mT
on a sampled waveform. In Section 2.7, we discussed the z-transform of a signal
delayed by T and advanced by mT, which is known as the modified z-transform.
To express (3.29) in terms of the modified z-transform, we divide the time delay lT
into a delay (l 2 1)T and a delay T and rewrite the transfer function as
GZAS ðzÞ 5 z2ðl21Þ ð1 2 z21 Þz21 Z fgs ðkT 1 mTÞg
(3.30)
Finally, we express the z-transfer function in terms of the modified z-transform
GZAS ðzÞ 5
z21
Z m gs ðkTÞg
l
z
(3.31)
We recall two important modified transforms that are given in Section 2.7:
1
z21
(3.32)
e2mpT
z 2 e2pT
(3.33)
Z m f1ðkTÞg 5
Z m fe2pkT g 5
Returning to our earlier numerical example, a delay of 3.1 sampling periods gives
a delay of 4 sampling periods and the corresponding z24 term and the modified
z-transform of g(kT) with m 5 0.9.
EXAMPLE 3.7
If the sampling period is 0.1 s, determine the z-transfer function GZAS(z) for the system
GðsÞ 5
3e20:31s
s13
Solution
First, write the delay in terms of the sampling period as 0.31 5 3.1 3 0.1 5 (4 2 0.9) 3 0.1.
Thus, l 5 4 and m 5 0.9. Next, obtain the partial fraction expansion
Gs ðsÞ 5
3
1
1
5 2
sðs 1 3Þ
s
s13
3.7 The closed-loop transfer function
gs(t)
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0
0.5
1
1.5
2
Time s
FIGURE 3.13
Continuous time function gs(t).
This is the transform of the continuous-time function shown in Figure 3.13,
which must be sampled, shifted, and z-transformed to obtain the desired transfer
function. Using the modified z-transforms obtained in Section 2.7, the desired transfer function is
9
18
< 1
20:3 3 0:9 =
z
2
1
e
GZAS ðzÞ 5 @ 4 A
2
:z 2 1
z 2 e20:3 ;
z
0
9
8
<z 2 0:741 2 0:763ðz 2 1Þ = 0:237z 1 0:022
5z
5 4
;
:
z 2 0:741
z ðz 2 0:741Þ
24
3.7 The closed-loop transfer function
Using the results of Section 3.5, the digital control system of Figure 3.1 yields the
closed-loop block diagram of Figure 3.14. The block diagram includes a comparator, a digital controller with transfer function C(z), and the ADC-analog subsystemDAC transfer function GZAS(z). The controller and comparator are actually
computer programs and replace the computer block in Figure 3.1. The block diagram is identical to those commonly encountered in s-domain analysis of analog
71
72
CHAPTER 3 Modeling of Digital Control Systems
R(z) +
E(z)
Y(z)
U(z)
GZAS(z)
C(z)
–
FIGURE 3.14
Block diagram of a single-loop digital control system.
systems, with the variable s replaced by z. Hence, the closed-loop transfer function
for the system is given by
Gcl ðzÞ 5
CðzÞGZAS ðzÞ
1 1 CðzÞGZAS ðzÞ
(3.34)
and the closed-loop characteristic equation is
1 1 CðzÞGZAS ðzÞ 5 0
(3.35)
The roots of the equation are the closed-loop system poles, which can be selected
for desired time response specifications as in s-domain design. Before we discuss
this in some detail, we first examine alternative system configurations and their
transfer functions.
When deriving closed-loop transfer functions of other configurations, the
results of Section 3.4 must be considered carefully, as seen from Example 3.8.
EXAMPLE 3.8
Find the Laplace transform of the analog and sampled output for the block diagram of
Figure 3.15.
Solution
The analog variable x(t) has the Laplace transform
XðsÞ 5 HðsÞGðsÞDðsÞEðsÞ
T
R(s) +
U(s)
E(s)
G(s)
D(s)
Y(s)
–
T
X *(s)
X(s)
H(s)
FIGURE 3.15
Block diagram of a system with sampling in the feedback path.
Y * (s)
3.7 The closed-loop transfer function
which involves three multiplications in the s-domain. In the time domain, x(t) is obtained
after three convolutions.
From the block diagram
EðsÞ 5 RðsÞ 2 X ðsÞ
Substituting in the X(s) expression, sampling then gives
XðsÞ 5 HðsÞGðsÞDðsÞ½RðsÞ 2 X ðsÞ
Thus, the impulse-sampled variable x (t) has the Laplace transform
X ðsÞ 5 ðHGDRÞ ðsÞ 2 ðHGDÞ ðsÞX ðsÞ
where, as in the first part of Example 3.1, several components are no longer separable.
These terms are obtained as shown in Example 3.1 by inverse Laplace transforming,
impulse sampling, and then Laplace transforming the impulse-sampled waveform.
Next, we solve for X (s)
X ðsÞ 5
ðHGDRÞ ðsÞ
1 1ðHGDÞ ðsÞ
and then E(s)
EðsÞ 5 RðsÞ 2
ðHGDRÞ ðsÞ
1 1ðHGDÞ ðsÞ
With some experience, the last two expressions can be obtained from the block diagram
directly. The combined terms are clearly the ones not separated by samplers in the block
diagram.
From the block diagram, the Laplace transform of the output is Y(s) 5 G(s)D(s)E(s).
Substituting for E(s) gives
YðsÞ 5 GðsÞDðsÞ RðsÞ 2
ðHGDRÞ ðsÞ
1 1ðHGDÞ ðsÞ
Thus, the sampled output is
Y ðsÞ 5 ðGDRÞ ðsÞ 2 ðGDÞ ðsÞ
ðHGDRÞ ðsÞ
1 1ðHGDÞ ðsÞ
With the transformation z 5 est, we can rewrite the sampled output as
YðzÞ 5 ðGDRÞðzÞ 2 ðGDÞðzÞ
ðHGDRÞðzÞ
1 1ðHGDÞðzÞ
The last equation demonstrates how for some digital systems, no expression
is available for the transfer function excluding the input. Nevertheless, the preceding system has a closed-loop characteristic equation similar to (3.35) given
by 1 1 (HGD)(z) 5 0. This equation can be used in design, as in cases where a
closed-loop transfer function is defined.
73
74
CHAPTER 3 Modeling of Digital Control Systems
3.8 Analog disturbances in a digital system
Disturbances are variables that are not included in the system model but affect
its response. They can be deterministic, such as load torque in a position control
system, or stochastic, such as sensor or actuator noise. However, almost all disturbances are analog and are inputs to the analog subsystem in a digital control loop.
We use the results of Section 3.7 to obtain a transfer function with a disturbance
input.
Consider the system with disturbance input shown in Figure 3.16. Because the
system is linear, the reference input can be treated separately and is assumed to
be zero.
The Laplace transform of the impulse-sampled output is
Y ðsÞ 5 ðGGd DÞ ðsÞ 2 ðGGZOH Þ ðsÞC ðsÞY ðsÞ
(3.36)
Solving for Y (s), we obtain
Y ðsÞ 5
ðGGd DÞ ðsÞ
1 1ðGGZOH Þ ðsÞC ðsÞ
(3.37)
The denominator involves the transfer function for the zero-order hold, analog
subsystem, and sampler. We can therefore rewrite (3.37) using the notation of
(3.21) as
Y ðsÞ 5
ðGGd DÞ ðsÞ
1 1ðGZAS Þ ðsÞC ðsÞ
(3.38)
ðGGd DÞðzÞ
1 1 GZAS ðzÞCðzÞ
(3.39)
or in terms of z as
Y ðzÞ 5
D(s)
Gd (s)
+
T
C*(s)
−
GZOH (s)
+
FIGURE 3.16
Block diagram of a digital system with an analog disturbance.
T
G(s)
Y *(s)
3.9 Steady-state error and error constants
EXAMPLE 3.9
Consider the block diagram of Figure 3.16 with the transfer functions
GðsÞ 5
Kp
;
s11
1
Gd ðsÞ 5 ;
s
CðzÞ 5 Kc
Find the steady-state response of the system to an impulse disturbance of strength A.
Solution
We first evaluate
GðsÞGd ðsÞDðsÞ 5
Kp A
1
1
5 Kp A 2
sðs 1 1Þ
s
s11
The z-transform of the corresponding impulse response sequence is
z
z
2
ðGGd DÞðzÞ 5 Kp A
z 2 1 z 2 e2T
Using (3.21), we obtain the transfer function
GZAS ðzÞ 5 Kp
1 2 e2T
z 2 e2T
From (3.39), we obtain the sampled output
z
z
2
z 2 1 z 2 e2T
YðzÞ 5
1 2 e2T
1 1 Kc Kp
2T
z2e
Kp A
To obtain the steady-state response, we use the final value theorem
yðNÞ 5 ðz21ÞYðzÞz51
Kp A
5
1 1 Kc Kp
Thus, as with analog systems, increasing the controller gain reduces the error due to the
disturbance. Equivalently, an analog amplifier before the point of disturbance injection
can increase the gain and reduce the output due to the disturbance and is less likely
to saturate the DAC. Note that it is simpler to apply the final value theorem without simplification because terms not involving (z 2 1) drop out.
3.9 Steady-state error and error constants
In this section, we consider the unity feedback block diagram shown in
Figure 3.14 subject to standard inputs and determine the associated tracking
error in each case. The standard inputs considered are the sampled step, the
sampled ramp, and the sampled parabolic. As with analog systems, an error
constant is associated with each input, and a type number can be defined for any
75
76
CHAPTER 3 Modeling of Digital Control Systems
system from which the nature of the error constant can be inferred. All results are
obtained by direct application of the final value theorem.
From Figure 3.14, the tracking error is given by
EðzÞ 5
5
RðzÞ
1 1 GZAS ðzÞCðzÞ
RðzÞ
1 1 LðzÞ
where L(z) denotes the loop gain of the system.
Applying the final value theorem yields the steady-state error
eðNÞ 5 ð1 2 z21 ÞEðzÞz51
ðz21ÞRðzÞ 5
zð11LðzÞÞ (3.40)
(3.41)
z51
The limit exists if all (z 2 1) terms in the denominator cancel. This depends on
the reference input as well as on the loop gain.
To examine the effect of the loop gain on the limit, rewrite it in the form
LðzÞ 5
NðzÞ
;
ðz21Þn DðzÞ
n$0
(3.42)
where N(z) and D(z) are numerator and denominator polynomials, respectively,
with no unity roots. The following definition plays an important role in determining the steady-state error of unity feedback systems.
DEFINITION 3.1: TYPE NUMBER
The type number of the system is the number of unity poles in the system z-transfer
function.
The loop gain of (3.42) has n poles at unity and is therefore type n. These
poles play the same role as poles at the origin for an s-domain transfer function
in determining the steady-state response of the system. Note that s-domain poles
at zero play the same role as z-domain poles at e0.
Substituting from (3.42) in the error expression (3.41) gives
ðz21Þn11 DðzÞRðzÞ eðNÞ 5
zðNðzÞ1ðz21Þn DðzÞÞ z51
ðz21Þn11 Dð1ÞRðzÞ 5
Nð1Þ1ðz21Þn Dð1Þ z51
Next, we examine the effect of the reference input on the steady-state error.
(3.43)
3.9 Steady-state error and error constants
3.9.1 Sampled step input
The z-transform of a sampled unit step input is
z
RðzÞ 5
z21
Substituting in (3.41) gives the steady-state error
1 eðNÞ 5
11LðzÞ z51
(3.44)
The steady-state error can also be written as
eðNÞ 5
1
1 1 Kp
(3.45)
where Kp is the position error constant given by
Kp 5 Lð1Þ
(3.46)
Examining (3.42) shows that Kp is finite for type 0 systems and infinite for
systems of type 1 or higher. Therefore, the steady-state error for a sampled unit
step input is
8
1
<
; n50
(3.47)
eðNÞ 5 1 1 Lð1Þ
:
0;
n$1
3.9.2 Sampled ramp input
The z-transform of a sampled unit ramp input is
RðzÞ 5
Tz
ðz21Þ2
Substituting in (3.41) gives the steady-state error
T
½z21½11LðzÞ z51
1
5
Kv
eðNÞ 5
(3.48)
where Kv is the velocity error constant. The velocity error constant is thus given by
1
(3.49)
Kv 5 ðz21ÞLðzÞ
T
z51
77
78
CHAPTER 3 Modeling of Digital Control Systems
From (3.49), the velocity error constant is zero for type 0 systems, finite
for type 1 systems, and infinite for type 2 or higher systems. The corresponding
steady-state error is
8
>
>
<
N;
n50
T
eðNÞ 5 ðz21ÞLðzÞ ; n 5 1
>
z51
>
:
0
n$2
(3.50)
Similarly, it can be shown that for a sampled parabolic input, an acceleration
error constant given by
Ka 5
1
2
ðz21Þ
LðzÞ
2
T
z51
(3.51)
can be defined, and the associated steady-state error is
8
N;
n#1
>
>
<
T2
; n52
eðNÞ 5
ðz21Þ2 LðzÞz51
>
>
:
0;
n$3
(3.52)
EXAMPLE 3.10
Find the steady-state position error for the digital position control system with unity feedback
and with the transfer functions
GZAS ðzÞ 5
Kðz 1 aÞ
;
ðz 2 1Þðz 2 bÞ
CðzÞ 5
Kc ðz 2 bÞ
;
z2c
0 , a; b; c , 1
1. For a sampled unit step input
2. For a sampled unit ramp input
Solution
The loop gain of the system is given by
LðzÞ 5 CðzÞGZAS ðzÞ 5
KKc ðz 1 aÞ
ðz 2 1Þðz 2 cÞ
The system is type 1. Therefore, it has zero steady-state error for a sampled step input and
a finite steady-state error for a sampled ramp input given by
eðNÞ 5
T
T
12c
5
KKc 1 1 a
ðz21ÞLðzÞz51
Clearly, the steady-state error is reduced by increasing the controller gain and is also
affected by the choice of controller pole and zero.
3.10
MATLAB commands
EXAMPLE 3.11
Find the steady-state error for the analog system
GðsÞ 5
K
s1a
a.0
1. For proportional analog control with a unit step input
2. For proportional digital control with a sampled unit step input
Solution
The transfer function of the system can be written as
GðsÞ 5
K=a
s=a 1 1
a.0
Thus, the position error constant for analog control is K/a, and the steady-state error is
eðNÞ 5
1
a
5
1 1 Kp
K 1a
For digital control, it can be shown that for sampling period T, the DAC-plant-ADC z-transfer
function is
GZAS ðzÞ 5
K 1 2 e2aT
a z 2 e2aT
Thus, the position error constant for digital control is
Kp 5 GZAS ðzÞz51 5 K=a
and the associated steady-state error is the same as that of the analog system with proportional control. In general, it can be shown that the steady-state error for the same control
strategy is identical for digital or analog implementation.
3.10 MATLAB commands
The transfer function for the ADC, analog subsystem, and DAC combination
can be easily obtained using MATLAB. Assume that the sampling period is equal
to 0.1 s and that the transfer function of the analog subsystem is G.
3.10.1 MATLAB
The MATLAB command to obtain a digital transfer function from an analog
transfer function is
.. g 5 tfðnum;denÞ
.. gd 5 c2dðg; 0:1; 'method'Þ
79
80
CHAPTER 3 Modeling of Digital Control Systems
where num is a vector containing the numerator coefficients of the analog
transfer function in descending order, and den is a similarly defined vector of
denominator coefficients. For example, the numerator polynomial (2s2 1 4s 1 3)
is entered as
.. num 5 ½2; 4; 3
The term “method” specifies the method used to obtain the digital transfer
function. For a system with a zero-order hold and sampler (DAC and ADC),
we use
.. gd 5 c2dðg; 0:1; 'zoh'Þ
For a first-order hold, we use
.. gd 5 c2dðg; 0:1; 'foh'Þ
Other options of MATLAB commands are available but are not relevant to the
material presented in this chapter.
For a system with a time delay, the discrete transfer function can be obtained
using the commands
gdelay 5 tfðnum; den; 'inputdelay';TdÞ % Delay 5 Td
gdelay d 5 c2dðgdelay; 0:1; 'method'Þ
3.10.2 Simulink
Simulink is a MATLAB toolbox that provides a graphical language for
simulating dynamic systems. The vocabulary of the language is a set of block
operations that the user can combine to create a dynamic system. The block
operations include continuous and discrete-time, linear and nonlinear, as well as
user-defined blocks containing MATLAB functions. Additional blocks called
sources provide a wide variety of inputs. Sink blocks provide a means of recording
the system response, and the response can then be sent to MATLAB for processing
or plotting.
This section provides a brief introduction to Simulink and its simulation of
discrete-time systems that we hope will motivate the reader to experiment with
this powerful simulation tool. In our view, this is the only way to become proficient in the use of Simulink.
To start Simulink, start MATLAB then type
.. simulink
3.10
MATLAB commands
This opens a window called the Simulink Library Browser, which offers the
user many blocks to choose from for simulation. To create a model, click on
File . New . Model
This opens a model window that serves as a board on which the user can assemble model blocks. The blocks are selected from the Simulink Library Browser as
needed.
We begin with a simple simulation of a source and sink. Click on Sources and
find a Step block, and then drag it to the model window. Then click on Sinks and
find a Scope and drag it to the model area. To connect two blocks, click on the first
block while holding down the Control button, and then click on the second block.
Simulink will automatically draw a line from the first block to the second. The
simulation diagram for this simple system is shown in Figure 3.17.
To set the parameters of a block, double-click it, and a window displaying the
parameters of the block will appear. Each of the parameters can be selected and
changed to the appropriate value for a particular simulation. For example, by
clicking on the Step block, we can select the time when the step is applied, the
input value before and after the step, and the sampling period. Double-clicking on
Scope shows a plot of the step after running the simulation. To run the simulation,
click on the arrow above the simulation diagram. The scope will show a plot of
the step input.
To model a discrete-time system, we select Discrete in the Simulink
Library Browser and drag a Discrete Transfer Function block to our model
window (see Figure 3.18). The numerator and denominator of the transfer
function are set by clicking on the block, and they are entered as MATLAB
vectors, as shown in Figure 3.19. The sampling period can also be selected
in the same window, and we select a value of 0.05 s. Starting with the SourceSink model, we can select the line joining the source to the sink and then click
on Delete to eliminate it. Then we can join the three blocks using the Control
button as before and save the new model. Running the simulation gives the
FIGURE 3.17
Simulation diagram for a step input and scope.
81
82
CHAPTER 3 Modeling of Digital Control Systems
FIGURE 3.18
Simulation diagram for a step input, discrete transfer function, and scope.
FIGURE 3.19
Parameters of the discrete transfer function.
results in Figure 3.20. The simulation duration is selected in the model window next to the run icon. The scale of the plot is selected by right-clicking on
the scope plot.
We now discuss simulating an analog system with digital control. The analog
transfer function is obtained by clicking on Continuous in the Simulink Library
Browser and then dragging the block to the model window. Although it is
not required to implement the zero-order hold, the zero-order hold block
is available under Discrete, but MATLAB does not have a sampler. The block
3.10
MATLAB commands
FIGURE 3.20
Step response of a discrete transfer function.
FIGURE 3.21
Simulation diagram for a step input, zero-order hold, analog transfer function,
and two scopes.
does not change the simulation results. However, the block allows the user to
obtain a plot of the input to the analog system.
The input to any discrete block is sampled automatically at the sampling rate
of the discrete block, and the sampling rate can also be chosen for sources and
sinks. Figure 3.21 shows the simulation diagram for the digital control system
with two scopes. The parameters of the scope are selected by clicking on the
parameter icon, which is located next to the printer (second from the left in
83
84
CHAPTER 3 Modeling of Digital Control Systems
FIGURE 3.22
Selecting the scope parameters.
FIGURE 3.23
Step response of analog system with digital control.
Problems
FIGURE 3.24
Step response of analog system with zero-order hold, without output sampling.
Figure 3.20). We set the sampling period for the first scope to 0.05 s and select
decimation and a value of 1 for the second scope to obtain the analog output on
the second scope, as shown in Figure 3.22. The sampled and analog outputs are
shown in Figures 3.23 and Figure 3.24, respectively.
Resources
Franklin, G.F., Powell, J.D., Workman, M.L., 1998. Digital Control of Dynamic Systems.
Addison-Wesley, Boston, MA.
Jacquot, R.G., 1981. Modern Digital Control Systems. Marcel Dekker, New York.
Katz, P., 1981. Digital Control Using Microprocessors. Prentice Hall International, Englewood
Cliffs, NJ.
Kuo, B.C., 1992. Digital Control Systems. Saunders, Ft. Worth, TX.
Ogata, K., 1987. Discrete-Time Control Systems. Prentice Hall, Englewood Cliffs, NJ.
PROBLEMS
3.1
Find the magnitude and phase at frequency ω 5 1 of a zero-order hold with
sampling time T 5 0.1 s.
85
86
CHAPTER 3 Modeling of Digital Control Systems
3.2
The first-order hold uses the last two numbers in a sequence to generate
its output. The output of both the zero-order hold and the first-order hold is
given by
uðtÞ 5 uðkTÞ 1 a
uðkTÞ 2 u½ðk 2 1ÞT
ðt 2 kTÞ; kT # t # ðk 1 1ÞT
T
k 5 0; 1; 2; . . .
with a 5 0,1, respectively.
(a) For a discrete impulse input, obtain and sketch the impulse response
for the preceding equation with a 5 0 and a 5 1.
(b) Write an impulse-sampled version of the preceding impulse response,
and show that the transfer function is given by
h
i
1
a
ð1 2 e2sT Þ
GH ðsÞ 5 ð1 2 e2sT Þ 1 2 ae2sT 1
s
sT
(c) Obtain the transfer functions for the zero-order hold and for the firstorder hold from the preceding transfer function. Verify that the transfer
function of the zero-order hold is the same as that obtained in
Section 3.3.
3.3
Many chemical processes can be modeled by the following transfer function:
GðsÞ 5
K
e2Td s
τs 1 1
where K is the gain, τ is the time constant, and Td is the time delay. Obtain
the transfer function GZAS(z) for the system in terms of the system parameters. Assume that the time delay Td is a multiple of the sampling time T.
3.4
Obtain the transfer function of a point mass (m) with force as input and displacement as output, neglecting actuator dynamics; then find GZAS(z) for
the system.
3.5
For an internal combustion engine, the transfer function with injected fuel
flow rate as input and fuel flow rate into the cylinder as output is given by1
GðsÞ 5
ετs 1 1
τs 1 1
where t is a time constant and e is known as the fuel split parameter.
Obtain the transfer function GZAS(z) for the system in terms of the system
parameters.
3.6
1
Repeat Problem 3.5 including a delay of 25 ms in the transfer function
with a sampling period of 10 ms.
Moskwa, J., 1988. Automotive Engine Modeling and Real time Control, MIT doctoral thesis.
Problems
3.7
Find the equivalent sampled impulse response sequence and the equivalent
z-transfer function for the cascade of the two analog systems with sampled
input
H1 ðsÞ 5
1
s16
H2 ðsÞ 5
10
s11
(a) If the systems are directly connected
(b) If the systems are separated by a sampler
3.8
Obtain expressions for the analog and sampled outputs from the block
diagrams shown in Figure P3.8.
R(s)
+
C(s)
−
U(s) U*(s)
Y * (s)
Y(s)
G(s)
T
T
H(s)
(a)
R(s)
+
E(s) E *(s)
−
G(s)
C(s)
Y * (s)
Y(s)
T
T
H(s)
(b)
R(s)
+
U(s) U*(s)
E(s) E* (s)
−
C(s)
Y *(s)
Y(s)
G(s)
T
T
T
H(s)
(c)
FIGURE P3.8
Block diagrams for systems with multiple samplers.
3.9
For the unity feedback system shown in Figure P3.9, we are given the analog
subsystem
GðsÞ 5
s18
s15
87
88
CHAPTER 3 Modeling of Digital Control Systems
R(z)
E(z)
+
U(z)
C(z)
−
Y(z)
DAC
ADC
G(s)
FIGURE P3.9
Block diagram for a closed-loop system with digital control.
The system is digitally controlled with a sampling period of 0.02 s. The
controller transfer function was selected as
CðzÞ 5
0:35z
z21
(a) Find the z-transfer function for the analog subsystem with DAC and ADC.
(b) Find the closed-loop transfer function and characteristic equation.
(c) Find the steady-state error for a sampled unit step and a sampled unit
ramp. Comment on the effect of the controller on steady-state error.
3.10
Find the steady-state error for a unit step disturbance input for the systems
shown in Figure P3.10 with a sampling period of 0.03 s and the transfer
functions
Gd ðsÞ 5
2
s11
GðsÞ 5
4ðs 1 2Þ
sðs 1 3Þ
C ðsÞ 5
esT 2 0:95
esT 2 1
D(s)
Gd (s)
R(s)
+
−
C* (s)
T
+
Y(s)
+
T
Y *(s)
G(s)
ZOH
T
(a)
D(s)
R(s)
+
− T
C* (s)
ZOH
Gd(s)
+
+
T
(b)
FIGURE P3.10
Block diagrams for systems with disturbance inputs.
Y(s) Y * (s)
G(s)
T
Computer Exercises
3.11
For the following systems with unity feedback, find
(a) The position error constant
(b) The velocity error constants
(c) The steady-state error for a unit step input
(d) The steady-state error for a unit ramp input
0:4ðz 1 0:2Þ
(i) GðzÞ 5
ðz 2 1Þðz 2 0:1Þ
0:5ðz 1 0:2Þ
(ii) GðzÞ 5
ðz 2 0:1Þðz 2 0:8Þ
COMPUTER EXERCISES
3.12
For the analog system with a sampling period of 0.05 s
GðsÞ 5
10ðs 1 2Þ
sðs 1 5Þ
(a) Obtain the transfer function for the system with sampled input and output.
(b) Obtain the transfer function for the system with DAC and ADC.
(c) Obtain the unit step response of the system with sampled output and
analog input.
(d) Obtain the poles of the systems in (a) and (b), and the output of (c),
and comment on the differences between them.
3.13
For the system of Problem 3.9
(a) Obtain the transfer function for the analog subsystem with DAC and ADC.
(b) Obtain the step response of the open-loop analog system and the
closed-loop digital control system and comment on the effect of the
controller on the time response.
(c) Obtain the frequency response of the digital control system, and verify
that 0.02 s is an acceptable choice of sampling period. Explain briefly
why the sampling period is chosen based on the closed-loop rather
than the open-loop dynamics.
3.14
Consider the internal combustion engine model of Problem 3.5. Assume
that for the operational conditions of interest, the time constant τ is approximately 1.2 s, whereas the parameter e can vary in the range 0.4 to 0.6. The
digital cascade controller
CðzÞ 5
0:02z
z21
was selected to improve the time response of the system with unity feedback. Simulate the digital control system with e 5 0.4, 0.5, and 0.6, and
discuss the behavior of the controller in each case.
89
90
CHAPTER 3 Modeling of Digital Control Systems
3.15
Simulate the continuous-discrete system discussed in Problem 3.9 and
examine the behavior of both the continuous output and the sampled
output. Repeat the simulation with a 10% error in the plant gain. Discuss
the simulation results, and comment on the effect of the parameter error on
disturbance rejection.
CHAPTER
Stability of Digital Control
Systems
4
OBJECTIVES
After completing this chapter, the reader will be able to do the following:
1. Determine the input-output stability of a z-transfer function.
2. Determine the asymptotic stability of a z-transfer function.
3. Determine the internal stability of a digital feedback control system.
4. Determine the stability of a z-polynomial using the Routh-Hurwitz criterion.
5. Determine the stability of a z-polynomial using the Jury criterion.
6. Determine the stable range of a parameter for a z-polynomial.
7. Determine the closed-loop stability of a digital system using the Nyquist criterion.
8. Determine the gain margin and phase margin of a digital system.
Stability is a basic requirement for digital and analog control systems. Digital
control is based on samples and is updated every sampling period, and there is a
possibility that the system will become unstable between updates. This obviously
makes stability analysis different in the digital case. We examine different definitions and tests of the stability of linear time-invariant (LTI) digital systems based
on transfer function models. In particular, we consider input-output stability and
internal stability. We provide several tests for stability: the Routh-Hurwitz criterion, the Jury criterion, and the Nyquist criterion. We also define the gain margin
and phase margin for digital systems.
4.1 Definitions of stability
The most commonly used definitions of stability are based on the magnitude
of the system response in the steady state. If the steady-state response is
unbounded, the system is said to be unstable. In this chapter, we discuss two
stability definitions that concern the boundedness or exponential decay of the
system output. The first stability definition considers the system output due to its
initial conditions. To apply it to transfer function models, we need the assumption
that no pole-zero cancellation occurs in the transfer function. Reasons for this
assumption are given in Chapter 8 and are discussed further in the context of
state-space models.
91
92
CHAPTER 4 Stability of Digital Control Systems
DEFINITION 4.1: ASYMPTOTIC STABILITY
A system is said to be asymptotically stable if its response to any initial conditions decays to
zero asymptotically in the steady state—that is, the response due to the initial conditions
satisfies
Lim yðkÞ 5 0
(4.1)
k-N
If the response due to the initial conditions remains bounded but does not decay to zero,
the system is said to be marginally stable.
The second definition of stability concerns the forced response of the system
for a bounded input. A bounded input satisfies the condition
uðkÞ , bu ;
k 5 0; 1; 2; . . .
0 , bu , N
(4.2)
For example, a bounded sequence satisfying the constraint ju(k)j , 3 is shown in
Figure 4.1.
u(k)
3
2
1
0
–1
–2
–3
0
0.5
1
1.5
2
FIGURE 4.1
Bounded sequence with bound bu 5 3.
2.5
3
3.5
4
4.5
5 kT
4.2 Stable z-domain pole locations
DEFINITION 4.2: BOUNDED-INPUTBOUNDED-OUTPUT STABILITY
A system is said to be bounded-inputbounded-output (BIBO) stable if its response to
any bounded input remains bounded—that is, for any input satisfying (4.2), the output
satisfies
yðkÞ , by ; k 5 0; 1; 2; . . .
(4.3)
0 , by , N
4.2 Stable z-domain pole locations
The examples provided in Chapter 2 show that the locations of the poles of the
system z-transfer function determine the time response. The implications of this
fact for system stability are now examined more closely.
Consider the sampled exponential and its z-transform
pk ;
k 5 0; 1; 2; . . .
Z
!
z
z2p
(4.4)
where p is real or complex. Then the time sequence for large k is given by
8
k < 0;
p - 1;
:
N;
p , 1
p 5 1
p . 1
(4.5)
Any time sequence can be described by
f ðkÞ 5
n
X
i51
Ai pki ;
k 5 0; 1; 2; . . .
Z
! F ðzÞ 5
n
X
i51
Ai
z
z 2 pi
(4.6)
where Ai are partial fraction coefficients and pi are z-domain poles. Hence,
we conclude that the sequence is bounded if its poles lie in the closed unit disc
(i.e., on or inside the unit circle) and decays exponentially if its poles lie in the
open unit disc (i.e., inside the unit circle). This conclusion allows us to derive stability conditions based on the locations of the system poles. Note that the case of
repeated poles on the unit circle corresponds to an unbounded time sequence (see,
for example, the transform of the sampled ramp).
Although the preceding conclusion is valid for complex as well as real-time
sequences, we will generally restrict our discussions to real-time sequences. For
real-time sequences, the poles and partial fraction coefficients in (4.6) are either
real or complex conjugate pairs.
93
94
CHAPTER 4 Stability of Digital Control Systems
4.3 Stability conditions
The analysis of section 4.2 allows the derivation of conditions for asymptotic and
BIBO stability based on transfer function models. It is shown that, in the absence
of pole-zero cancellation, conditions for BIBO stability and asymptotic stability
are identical.
4.3.1 Asymptotic stability
Theorem 4.1 gives conditions for asymptotic stability.
THEOREM 4.1: ASYMPTOTIC STABILITY
In the absence of pole-zero cancellation, an LTI digital system is asymptotically stable if its
transfer function poles are in the open unit disc and marginally stable if the poles are in the
closed unit disc with no repeated poles on the unit circle.
PROOF
Consider the LTI system governed by the constant coefficient difference equation
yðk 1 nÞ 1 an21 yðk 1 n 2 1Þ 1 . . . 1 a1 yðk 1 1Þ 1 a0 yðkÞ
5 bm uðk 1 mÞ 1 bm21 uðk 1 m 2 1Þ 1 . . . 1 b1 uðk 1 1Þ 1 b0 uðkÞ;
k 5 0; 1; 2; . . .
with initial conditions y(0), y(1), y(2), . . . , y(n 2 1). Using the z-transform of the output, we
observe that the response of the system due to the initial conditions with the input zero is of
the form
Y ðzÞ 5
N ðzÞ
zn 1 an21 zn21 1 . . . 1 a1 z 1 a0
where N(z) is a polynomial dependent on the initial conditions. Because transfer function
zeros arise from transforming the input terms, they have no influence on the response due to
the initial conditions. The denominator of the output z-transform is the same as the denominator of the z-transfer function in the absence of pole-zero cancellation. Hence, the poles
of the function Y(z) are the poles of the system transfer function.
Y(z) can be expanded as partial fractions of the form (4.6). Thus, the output due to the
initial conditions is bounded for system poles in the closed unit disc with no repeated poles
on the unit circle. It decays exponentially for system poles in the open unit disc (i.e., inside
the unit circle).
Theorem 4.1 applies even if pole-zero cancellation occurs, provided that the
poles that cancel are stable. This follows from the fact that stability testing is
essentially a search for unstable poles, with the system being declared stable
if none are found. Invisible but stable poles would not lead us to a wrong conclusion. However, invisible but unstable poles would. The next example shows how
to determine asymptotic stability using Theorem 4.1.
4.3 Stability conditions
EXAMPLE 4.1
Determine the asymptotic stability of the following systems:
(a) HðzÞ 5
4ðz 2 2Þ
ðz 2 2Þðz 2 0:1Þ
(b) HðzÞ 5
4ðz 2 0:2Þ
ðz 2 0:2Þðz 2 0:1Þ
(c) HðzÞ 5
5ðz 2 0:3Þ
ðz 2 0:2Þðz 2 0:1Þ
(d) HðzÞ 5
8ðz 2 0:2Þ
ðz 2 0:1Þðz 2 1Þ
Solution
Theorem 4.1 can only be used for transfer functions (a) and (b) if their poles and zeros are
not canceled. Ignoring the zeros, which do not affect the response to the initial conditions,
(a) has a pole outside the unit circle and the poles of (b) are inside the unit circle. Hence,
(a) is unstable, whereas (b) is asymptotically stable.
Theorem 4.1 can be applied to the transfer functions (c) and (d). The poles of (c) are
all inside the unit circle, and the system is therefore asymptotically stable. However, (d)
has one pole on the unit circle and is only marginally stable.
4.3.2 BIBO stability
BIBO stability concerns the response of a system to a bounded input. The response
of the system to any input is given by the convolution summation
yðkÞ 5
k
X
hðk 2 iÞuðiÞ;
k 5 0; 1; 2; . . .
(4.7)
i50
where h(k) is the impulse response sequence.
It may seem that a system should be BIBO stable if its impulse response is
bounded. To show that this is generally false, let the impulse response of a linear
system be bounded and strictly positive with lower bound bh1 and upper bound
bh2—that is,
0 , bh1 , hðkÞ , bh2 , N;
k 5 0; 1; 2; . . .
(4.8)
Then using the bound (4.8) in (4.7) gives the inequality
k
k
X
X
yðkÞ 5
hðk 2 iÞuðiÞ . bh1
uðiÞ;
i50
k 5 0; 1; 2; . . .
(4.9)
i50
which is unbounded as k-N for the bounded input u(k) 5 1, k 5 0, 1, 2, . . .
Theorem 4.2 establishes necessary and sufficient conditions for BIBO stability
of a discrete-time linear system.
95
96
CHAPTER 4 Stability of Digital Control Systems
THEOREM 4.2
A discrete-time linear system is BIBO stable if and only if its impulse response sequence is
absolutely summable—that is,
N X
hðiÞ , N
(4.10)
i50
PROOF
1. Necessity (only if)
To prove necessity by contradiction, assume that the system is BIBO stable but does not
satisfy (4.10). Then consider the input sequence given by
1;
hðiÞ $ 0
uðk 2 iÞ 5
21; hðiÞ , 0
The corresponding output is
yðkÞ 5
k X
hðiÞ
i50
which is unbounded as k-N. This contradicts the assumption of BIBO stability.
2. Sufficiency (if)
To prove sufficiency, we assume that (4.10) is satisfied and then show that the system
is BIBO stable. Using the bound (4.2) in the convolution summation (4.7) gives the
inequality
k k X
X
yðkÞ #
hðiÞuðk 2 iÞ , bu
hðiÞ;
i50
k 5 0; 1; 2; . . .
i50
which remains bounded as k-N if (4.10) is satisfied.
Because the z-transform of the impulse response is the transfer function,
BIBO stability can be related to pole locations as shown in Theorem 4.3.
THEOREM 4.3
A discrete-time linear system is BIBO stable if and only if the poles of its transfer function
lie inside the unit circle.
PROOF
Applying (4.6) to the impulse response and transfer function shows that the impulse
response is bounded if the poles of the transfer function are in the closed unit disc and
decays exponentially if the poles are in the open unit disc. It has already been established
that systems with a bounded impulse response that does not decay exponentially are not
BIBO stable. Thus, for BIBO stability, the system poles must lie inside the unit circle.
4.3 Stability conditions
To prove sufficiency, assume an exponentially decaying impulse response (i.e., poles
inside the unit circle). Let Ar be the coefficient of largest magnitude and jpsj , 1 be the
system pole of largest magnitude in (4.6). Then the impulse response (assuming no repeated
poles for simplicity) is bounded by
X
n
n
X
k
k
hðkÞ 5 jA jp # njAr jps ;
Ai pki #
k 5 0; 1; 2; . . .
i51 i i
i51
Hence, the impulse response decays exponentially at a rate determined by the largest system
pole. Substituting the upper bound in (4.10) gives
N N X
X
hðiÞ # njAr j
ps i 5 njAr j
i50
i50
1
,N
1 2 ps Thus, the condition is sufficient by Theorem 4.2.
EXAMPLE 4.2
Investigate the BIBO stability of the class of systems with the impulse response
hðkÞ 5
K; 0 # k # m , N
0; elsewhere
where K is a finite constant.
Solution
The impulse response satisfies
N m X
X
hðiÞ 5
hðiÞ 5 ðm 1 1ÞjK j , N
i50
i50
Using condition (4.10), the systems are all BIBO stable. This is the class of finite impulse
response (FIR) systems (i.e., systems whose impulse response is nonzero over a finite interval).
Thus, we conclude that all FIR systems are BIBO stable.
EXAMPLE 4.3
Investigate the BIBO stability of the systems discussed in Example 4.1.
Solution
After pole-zero cancellation, the transfer functions (a) and (b) have all poles inside the
unit circle and are therefore BIBO stable. The transfer function (c) has all poles inside the
unit circle and is stable; (d) has a pole on the unit circle and is not BIBO stable.
The preceding analysis and examples show that for LTI systems, with no
pole-zero cancellation, BIBO and asymptotic stability are equivalent and can be
investigated using the same tests. Hence, the term stability is used in the sequel
to denote either BIBO or asymptotic stability with the assumption of no unstable polezero cancellation. Pole locations for a stable system (inside the unit circle) are shown
in Figure 4.2.
97
98
CHAPTER 4 Stability of Digital Control Systems
Im[z]
Unit Circle
Stable
Re[z]
Unstable
FIGURE 4.2
Stable pole locations in the z-plane. Im[z] denotes the imaginary part and Re[z] denotes
the real part of z.
D(z)
R(z) +
E(z)
C(z)
U(z)
Y(z)
GZAS (z)
−
FIGURE 4.3
Digital control system with disturbance D(z).
4.3.3 Internal stability
So far, we have only considered stability as applied to an open-loop system.
For closed-loop systems, these results are applicable to the closed-loop transfer
function. However, the stability of the closed-loop transfer function is not always
sufficient for proper system operation because some of the internal variables may
be unbounded. In a feedback control system, it is essential that all the signals in
the loop be bounded when bounded exogenous inputs are applied to the system.
Consider the unity feedback digital control scheme of Figure 4.3 where, for
simplicity, a disturbances input is added to the controller output before the ADC.
We consider that system as having two outputs, Y and U, and two inputs, R and D.
Thus, the transfer functions associated with the system are given by
2
3
CðzÞGZAS ðzÞ
GZAS ðzÞ
6 1 1 CðzÞG ðzÞ
1 1 CðzÞGZAS ðzÞ 7
ZAS
Y ðzÞ
6
7 RðzÞ
(4.11)
56
CðzÞ
CðzÞGZAS ðzÞ 7
U ðzÞ
4
5 DðzÞ
2
1 1 CðzÞGZAS ðzÞ
1 1 CðzÞGZAS ðzÞ
Clearly, it is not sufficient to prove that the output of the controlled system Y
is bounded for bounded reference input R because the controller output U can be
unbounded. In addition, the system output must be bounded when a different
4.3 Stability conditions
input is applied to the system—namely, in the presence of a disturbance. This
suggests the following definition of stability.
DEFINITION 4.3: INTERNAL STABILITY
If all the transfer functions that relate system inputs (R and D) to the possible system
outputs (Y and U) are BIBO stable, then the system is said to be internally stable.
Because internal stability guarantees the stability of the transfer function from R
to Y, among others, it is obvious that an internally stable system is also externally
stable (i.e., the system output Y is bounded when the reference input R is bounded).
However, external stability does not, in general, imply internal stability.
We now provide some results that allow us to test internal stability.
THEOREM 4.4
The system shown in Figure 4.3 is internally stable if and only if all its closed-loop poles are
in the open unit disc.
PROOF
1. Necessity (only if)
To prove necessity, we write C(z) and GZAS(z) as ratios of coprime polynomials (i.e., polynomials with no common factors):
CðzÞ 5
NC ðzÞ
DC ðzÞ
GZAS ðzÞ 5
Substituting in (4.11), we rewrite it as
1
Y
NC NG
5
U
DC DG 1 NC NG NC DG
NG ðzÞ
DG ðzÞ
DC NG
2NC NG
(4.12)
R
D
(4.13)
where we have dropped the argument z for brevity. If the system is internally stable, then the
four transfer functions in (4.11) have no poles on or outside the unit circle. Because of the
factorization of (4.12) is coprime, the polynomial DCDG 1 NCNG has no zeros that cancel
with all four numerators. Thus, the polynomial has no zeros on or outside the unit circle.
2. Sufficiency (if)
From (4.13), it is evident that if the characteristic polynomial DCDG 1 NCNG has no zeros
on or outside the unit circle, then all the transfer functions are asymptotically stable and
the system is internally stable.
THEOREM 4.5
The system in Figure 4.3 is internally stable if and only if the following two conditions hold:
1. The characteristic polynomial 1 1 C(z)GZAS(z) has no zeros on or outside the unit circle.
2. The loop gain C(z)GZAS(z) has no pole-zero cancellation on or outside the unit circle.
99
100
CHAPTER 4 Stability of Digital Control Systems
PROOF
1. Necessity (only if)
Condition 1 is clearly necessary by Theorem 4.4. To prove the necessity of condition 2, we
first factor C(z) and GZAS(z) as in (4.12) to write the characteristic polynomial in the form
DCDG 1 NCNG. We also have that C(z) GZAS(z) is equal to NCNG/DCDG. Assume that condition 2 is violated and that there exists Z0, jZ0j $ 1, which is a zero of DCDG as well as a zero
of NCNG. Then clearly Z0 is also a zero of the characteristic polynomial DCDG 1 NCNG, and
the system is unstable. This establishes the necessity of condition 2.
2. Sufficiency (if)
By Theorem 4.4, condition 1 implies internal stability unless unstable pole-zero cancellation occurs in the characteristic polynomial 1 1 C(z)GZAS(z). We therefore have internal
stability if condition 2 implies the absence of unstable pole-zero cancellation. If the
loop gain C(z)GZAS(z) 5 NCNG/DCDG has no unstable pole-zero cancellation, then 1 1 C(z)
GZAS(z) 5 [DCDG 1 NCNG]/DCDG does not have unstable pole-zero cancellation, and the
system is internally stable.
EXAMPLE 4.4
An isothermal chemical reactor where the product concentration is controlled by manipulating the feed flow rate is modeled by the following transfer function:1
GðsÞ 5
0:5848ð20:3549s 1 1Þ
0:1828s2 1 0:8627 1 1
Determine GZAS(Z) with a sampling rate T 5 0.1, and then verify that the closed-loop
system with the feedback controller
CðzÞ 5
210ðz 2 0:8149Þðz 2 0:7655Þ
ðz 2 1Þðz 2 1:334Þ
is not internally stable.
Solution
The discretized process transfer function is
GZAS ðzÞ 5 ð1 2 z21 Þ Z
GðsÞ
20:075997ðz 2 1:334Þ
5
s
ðz 2 0:8149Þðz 2 0:7655Þ
The transfer function from the reference input to the system output is given by
YðzÞ
CðzÞGZAS ðzÞ
0:75997
5
5
RðzÞ 1 1 CðzÞGZAS ðzÞ z 2 0:24
The system appears to be asymptotically stable with all its poles inside the unit circle.
However, the system is not internally stable as seen by examining the transfer function
UðzÞ
CðzÞ
210ðz 2 0:8149Þðz 2 0:7655Þ
5
5
RðzÞ
1 1 CðzÞGZAS ðzÞ
ðz 2 0:24Þðz 2 1:334Þ
4.4 Stability determination
which has a pole at 1.334 outside the unit circle. The control variable is unbounded even
when the reference input is bounded. In fact, the system violates condition 2 of Theorem 4.5
because the pole at 1.334 cancels in the loop gain
CðzÞGZAS ðzÞ 5
1
210ðz 2 0:8149Þðz 2 0:7655Þ
20:075997ðz 2 1:334Þ
3
ðz 2 1Þðz 2 1:334Þ
ðz 2 0:8149Þðz 2 0:7655Þ
Bequette, B. W., 2003. Process Control: Modeling, Design, and Simulation, Prentice Hall, Upper Saddle River, NJ.
4.4 Stability determination
The simplest method for determining the stability of a discrete-time system given
its z-transfer function is by finding the system poles. This can be accomplished
using a suitable numerical algorithm based on Newton’s method.
4.4.1 MATLAB
The roots of a polynomial are obtained using one of the MATLAB commands
.. rootsðdenÞ
.. zpkðgÞ
where den is a vector of denominator polynomial coefficients. The command zpk
factorizes the numerator and denominator of the transfer function g and displays it.
The poles of the transfer function can be obtained with the command pole and then
sorted with the command dsort in order of decreasing magnitude.
Alternatively, one may use the command ddamp, which yields the pole
locations (eigenvalues), the damping ratio, and the undamped natural frequency.
For example, given a sampling period of 0.1 s and the denominator polynomial
with coefficients
.. den 5 ½1:0; 0:2; 0:0; 0:4
the command is
.. ddampðden;0:1Þ
The command yields the output.
Eigenvalue
Magnitude
Equiv. Damping
Equiv. Freq. (rad/sec)
0.2306 1 0.7428I
0.2306 2 0.7428I
20.6612
0.7778
0.7778
0.6612
0.1941
0.1941
0.1306
12.9441
12.9441
31.6871
101
102
CHAPTER 4 Stability of Digital Control Systems
The MATLAB command
.. T 5 feedbackðg;gf; 61Þ
calculates the closed-loop transfer function T using the forward transfer function
g and the feedback transfer function gf. For negative feedback, the third argument
is 21 or is omitted. For unity feedback, we replace the argument gf with 1. We
can solve for the poles of the closed-loop transfer function as before using zpk or
ddamp.
4.4.2 Routh-Hurwitz criterion
The Routh-Hurwitz criterion determines conditions for left half plane (LHP) polynomial roots and cannot be directly used to investigate the stability of discrete-time
systems. The bilinear transformation
z5
11w
z21
3w 5
12w
z11
(4.14)
transforms the inside of the unit circle to the LHP.
To verify this property, consider the three cases shown in Figure 4.4. They
represent the mapping of a point in the LHP, a point in the RHP, and a point on
the jw axis. The angle of w after bilinear transformation is
+w 5 +ðz 2 1Þ 2 +ðz 1 1Þ
(4.15)
For a point inside the unit circle, the angle of w is of a magnitude greater
than 90 , which corresponds to points in the LHP. For a point on the unit circle,
the angle is 690 , which corresponds to points on the imaginary axis, and
for points outside the unit circle, the magnitude of the angle is less than 90 , which
corresponds to points in the RHP.
Im(w)
Im(z)
z3
X w3
z3 − 1
z2 + 1
z3 + 1
z2
w2 X
Re(z)
Re(w)
1
w1 X
FIGURE 4.4
Angles associated with bilinear transformation.
z2 − 1
z1 + 1
−1
z1
z1 − 1
4.4 Stability determination
The bilinear transformation allows the use of the Routh-Hurwitz criterion for
the investigation of discrete-time system stability. For the general z-polynomial,
we have the transform pairs
w
z 5 11 1
11w n
11w n21 . . .
2w
n
n21
.
.
.
1 an21
1 1 a0
FðzÞ 5 an z 1 an21 z 1 1 a0 ! an
12w
12w
(4.16)
The Routh-Hurwitz approach becomes progressively more difficult as the
order of the z-polynomial increases. But for low-order polynomials, it easily gives
stability conditions. For high-order polynomials, a symbolic manipulation package
can be used to perform the necessary algebraic manipulations. The RouthHurwitz approach is demonstrated in Example 4.5.
EXAMPLE 4.5
Find stability conditions for
1. The first-order polynomial a1z 1 a0, a1 . 0
2. The second-order polynomial a2z2 1 a1z 1 a0, a2 . 0
Solution
1. The stability of the first-order polynomial can be easily determined by solving for its
root. Hence, the stability condition is
a0 ,1
(4.17)
a 1
2. The roots of the second-order polynomial are in general given by
z1;2 5
2a1 6
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
a21 2 4a0 a2
2a2
(4.18)
Thus, it is not easy to determine the stability of the second-order polynomial by solving for
its roots. For a monic polynomial (coefficient of z2 is unity), the constant term is equal to the
product of the poles. Hence, for pole magnitudes less than unity, we obtain the necessary
stability condition
a0 ,1
(4.19)
a 2
or equivalently
2a0 , a2
and
a0 , a2
This condition is also sufficient in the case of complex conjugate poles where the two
poles are of equal magnitude. The condition is only necessary for real poles because the
product of a number greater than unity and a number less than unity can be less than
unity. For example, for poles at 0.01 and 10, the product of the two poles has magnitude
0.1, which satisfies (4.19), but the system is clearly unstable.
Substituting the bilinear transformation in the second-order polynomial gives
a2
11w 2
11w
1 a1
1 a0
12w
12w
103
104
CHAPTER 4 Stability of Digital Control Systems
a0
1
Stable
−1
a1
1
−1
FIGURE 4.5
Stable parameter range for a second-order z-polynomial.
which reduces to
ða2 2 a1 1 a0 Þw2 1 2ða2 2 a0 Þw 1ða2 1 a1 1 a0 Þ
By the Routh-Hurwitz criterion, it can be shown that the poles of the second-order
w-polynomial remain in the LHP if and only if its coefficients are all positive. Hence, the
stability conditions are given by
a2 2 a1 1 a0 . 0
a2 2 a0 . 0
a2 1 a1 1 a0 . 0
(4.20)
Adding the first and third conditions gives
a2 1 a0 . 0. 2a0 , a2
This condition, obtained earlier in (4.19), is therefore satisfied if the three conditions of
(4.20) are satisfied. The reader can verify through numerical examples that if real roots satisfying conditions (4.20) are substituted in (4.18), we obtain roots between 21 and 11.
Without loss of generality, the coefficient a2 can be assumed to be unity, and the
stable parameter range can be depicted in the a0 versus a1 parameter plane as shown
in Figure 4.5.
4.5 Jury test
It is possible to investigate the stability of z-domain polynomials directly using the
Jury test for real coefficients or the Schur-Cohn test for complex coefficients.
These tests involve determinant evaluations as in the Routh-Hurwitz test for s-domain
polynomials but are more time consuming. The Jury test is given next.
THEOREM 4.6
For the polynomial
FðzÞ 5 an zn 1 an21 zn21 1 . . . 1 a1 z 1 a0 5 0;
an . 0
(4.21)
4.5 Jury test
Table 4.1 Jury Table
Row
z0
z1
z2
...
zn2k
...
zn21
zn
1
2
3
4
5
6
.
.
.
2 n25
2 n24
2 n23
a0
an
b0
bn21
c0
cn22
.
.
.
s0
s3
r0
a1
an21
b1
bn22
c1
cn23
.
.
.
s1
s2
r1
a2
an22
b2
bn23
c2
cn24
.
.
.
s2
s1
r2
...
...
...
...
...
...
...
...
...
s3
s0
an2k
ak
bn2k
bk
...
...
...
...
...
...
...
...
...
cn22
c0
an21
a1
bn21
b0
an
a0
the roots of the polynomial are inside the unit circle if and only if
ð1Þ
ð2Þ
ð3Þ
ð4Þ
ð5Þ
Fð1Þ . 0
ð21Þn Fð21Þ . 0
ja0 j , an
jb0 j . jbn21 j
jc0 j . jcn22 j
:
:
:
ðn 1 1Þ jr0 j . jr2 j
(4.22)
where the terms in the n 1 1 conditions are calculated from Table 4.1.
The entries of the table are calculated as follows:
a0 an2k ;
k 5 0; 1; . . . ; n 2 1
bk 5 a
ak n
b0
bn2k21 ; k 5 0; 1; . . . ; n 2 2
ck 5 bn21
bk :
:
:
s0 s3 ; r1 5 s0 s2 ; r2 5 s0 s1 r0 5 s s s s s s 3
0
3
1
3
(4.23)
2
Based on the Jury table and the Jury stability conditions, we make the following
observations:
1. The first row of the Jury table is a listing of the coefficients of the polynomial
F(z) in order of increasing power of z.
105
106
CHAPTER 4 Stability of Digital Control Systems
2. The number of rows of the table 2 n 2 3 is always odd, and the coefficients of
each even row are the same as the odd row directly above it with the order of
the coefficients reversed.
3. There are n 1 1 conditions in (4.22) that correspond to the n 1 1 coefficients
of F(z).
4. Conditions 3 through n 1 1 of (4.22) are calculated using the coefficient of the
first column of the Jury table together with the last coefficient of the
preceding row. The middle coefficient of the last row is never used and need
not be calculated.
5. Conditions 1 and 2 of (4.22) are calculated from F(z) directly. If one of the
first two conditions is violated, we conclude that F(z) has roots on or outside
the unit circle without the need to construct the Jury table or test the
remaining conditions.
6. Condition 3 of (4.22), with an 5 1, requires the constant term of the
polynomial to be less than unity in magnitude. The constant term is simply
the product of the roots and must be smaller than unity for all the roots to be
inside the unit circle.
7. Conditions (4.22) reduce to conditions (4.19) and (4.20) for first and
second-order systems, respectively, where the Jury table is simply one row.
8. For higher-order systems, applying the Jury test by hand is laborious, and it is
preferable to test the stability of a polynomial F(z) using a computer-aided
design (CAD) package.
9. If the coefficients of the polynomial are functions of system parameters,
the Jury test can be used to obtain the stable ranges of the system parameters.
EXAMPLE 4.6
Test the stability of the polynomial.
FðzÞ 5 z5 1 2:6z4 2 0:56z3 2 2:05z2 1 0:0775z 1 0:35 5 0
We compute the entries of the Jury table using the coefficients of the polynomial
(Table 4.2).
Table 4.2 Jury Table for Example 4.6
Row
z0
z1
z2
z3
z4
z5
1
2
3
4
5
6
7
0.35
1
20.8775
0.8325
0.0770
0.5151
20.2593
0.0775
2.6
22.5729
1.854
0.7143
0.2693
20.0837
22.05
20.56
20.1575
20.1575
0.2693
0.7143
20.3472
20.56
22.05
1.854
22.5729
0.5151
0.0770
2.6
0.0775
0.8325
20.8775
1
0.35
4.5 Jury test
The first two conditions require the evaluation of F(z) at z 5 61.
1. F(1) 5 1 1 2.6 2 0.56 2 2.05 1 0.0775 2 0.35 5 1.4175 . 0
2. (21)5F(21) 5 (21)(21 1 2.6 1 0.56 2 2.05 2 0.0775 1 0.35) 5 20.3825 , 0
Conditions 3 through 6 can be checked quickly using the entries of the first column of
the Jury table.
3. j0.35j , 1
4. j20.8775j . j0.8325j
5. j0.0770j , j0.5151j
6. j20.2593j , j20.3472j
Conditions 2, 5, and 6 are violated, and the polynomial has roots on or outside the unit
circle. In fact, the polynomial can be factored as
FðzÞ 5 ðz 2 0:7Þðz 2 0:5Þðz 1 0:5Þðz 1 0:8Þðz 1 2:5Þ 5 0
and has a root at 22.5 outside the unit circle. Note that the number of conditions violated
is not equal to the number of roots outside the unit circle and that condition 2 is sufficient
to conclude the instability of F(z).
EXAMPLE 4.7
Find the stable range of the gain K for the unity feedback digital cruise control system of
Example 3.2 with the analog plant transfer function
GðsÞ 5
K
s13
and with digital-to-analog converter (DAC) and analog-to-digital converter (ADC) if the
sampling period is 0.02 s.
Solution
The transfer function for analog subsystem ADC and DAC is
8
2
39
=
<
GðsÞ
5
GZAS ðzÞ 5 ð1 2 z21 Þ Z L 21 4
:
s ;
8
2
39
<
K 5=
21 4
5 ð1 2 z Þ Z L
:
sðs 1 3Þ ;
21
Using the partial fraction expansion
K
K 1
1
5
2
sðs 1 3Þ
3 s
s13
we obtain the transfer function
GZAS ðzÞ 5
1:9412 3 1022 K
z 2 0:9418
For unity feedback, the closed-loop characteristic equation is
1 1 GZAS ðzÞ 5 0
107
108
CHAPTER 4 Stability of Digital Control Systems
which can be simplified to
z 2 0:9418 1 1:9412 3 1022
K 50
The stability conditions are
0:9418 2 1:9412 3 1022 K , 1
20:9418 1 1:9412 3 1022 K , 1
Thus, the stable range of K is
23 , K , 100:03
EXAMPLE 4.8
Find the stable range of the gain K for the vehicle position control system (see Example 3.3)
with the analog plant transfer function
GðsÞ 5
10K
sðs 1 10Þ
and with DAC and ADC if the sampling period is 0.05 s.
Solution
The transfer function for analog subsystem ADC and DAC is
8
2
39
=
<
GðsÞ
21
5
GZAS ðzÞ 5 ð12 z21 Þ L 4
:
s ;
9
10K =
5 ð12 z Þ 2
:s ðs 1 10Þ ;
21
8
<
Using the partial fraction expansion
10K
10 1
1
5
0:1
K
2
1
s2 ðs 1 10Þ
s2
s
s 1 10
we obtain the transfer function
GZAS ðzÞ 5
1:0653 3 1022 Kðz 1 0:8467Þ
ðz 2 1Þðz 2 0:6065Þ
For unity feedback, the closed-loop characteristic equation is
1 1 GZAS ðzÞ 5 0
which can be simplified to
ðz 2 1Þðz 2 0:6065Þ 1 1:0653 3 1022 Kðz 1 0:8467Þ
5 z2 1ð1:0653 3 1022 K 2 1:6065Þz 1 0:6065 1 9:02 3 1023 K 5 0
The stability conditions are
1. F(1) 5 1 1 (1.0653 3 1022 K 2 1.6065) 1 0.6065 1 9.02 3 1023 K . 03K . 0
2. F(21) 5 1 2 (1.0653 3 1022 K 2 1.6065) 1 0.6065 1 9.02 3 1023 K . 03K , 1967.582
4.6 Nyquist criterion
3. j0.6065 1 0.0902 Kj , 13 1 (0.6065 1 0.0902 K) , 1
2 (0.6065 1 0.0902 K) , 1
3 2178.104 , K , 43.6199
The three conditions yield the stable range
0 , K , 43:6199
4.6 Nyquist criterion
The Nyquist criterion allows us to answer two questions:
1. Does the system have closed-loop poles outside the unit circle?
2. If the answer to the first question is yes, how many closed-loop poles are
outside the unit circle?
We begin by considering the closed-loop characteristic polynomial
pcl ðzÞ 5 1 1 CðzÞGðzÞ 5 1 1 LðzÞ 5 0
(4.24)
where L(z) denotes the loop gain. We rewrite the characteristic polynomial in
terms of the numerator of the loop gain NL and its denominator DL in the form
pcl ðzÞ 5 1 1
NL ðzÞ
NL ðzÞ 1 DL ðzÞ
5
DL ðzÞ
DL ðzÞ
(4.25)
We observe that the zeros of the rational function are the closed-loop poles,
whereas its poles are the open-loop poles of the system. We assume that we are
given the number of open-loop poles outside the unit circle, and we denote this
number by P. The number of closed-loop poles outside the unit circle is denoted
by Z and is unknown.
To determine Z, we use some simple results from complex analysis. We first
give the following definition.
DEFINITION 4.4: CONTOUR
A contour is a closed directed simple (does not cross itself) curve.
An example of a contour is shown in Figure 4.6. In the figure, shaded text
denotes a vector. Recall that in the complex plane the vector connecting any point
a to a point z is the vector (z 2 a). We can calculate the net angle change for the
term (z 2 a) as the point z traverses the contour in the shown (clockwise) direction by determining the net number of rotations of the corresponding vector.
From Figure 4.6, we observe that the net rotation is one full turn or 360 for the
point a1, which is inside the contour. The net rotation is zero for the point a2,
which is outside the contour. If the point in question corresponds to a zero, then
109
110
CHAPTER 4 Stability of Digital Control Systems
a1
z2
− a1
z1−a1
×
a2
z1
0
z1
FIGURE 4.6
Closed contours (shaded letters denote vectors).
the rotation gives a numerator angle; if it is a pole, we have a denominator angle.
The net angle change for a rational function is the change in the angle of the
numerator minus the change in the angle of the denominator. So for Figure 4.6,
we have one clockwise rotation because of a1 and no rotation as a result of a2 for
a net angle change of one clockwise rotation. Angles are typically measured
in the counterclockwise direction. We therefore count clockwise rotations as
negative.
The preceding discussion shows how to determine the number of zeros of a
rational function in a specific region; given the number of poles, we perform the
following steps:
1. Select a closed contour surrounding the region.
2. Compute the net angle change for the function as we traverse the contour
once.
3. The net angle change or number of rotations N is equal to the number of zeros
inside the contour Z minus the angle of poles inside the contour P. Calculate
the number of zeros as
Z 5N 1P
(4.26)
To use this to determine closed-loop stability, we need to select a contour that
encircles the outside of the unit circle. The contour is shown in Figure 4.7. The
smaller circle is the unit circle, whereas the large circle is selected with a large
radius so as to enclose all poles and zeros of the functions of interest.
The value of the loop gain on the unit circle is L(ejωT), which is the frequency
response of the discrete-time system for angles ωT in the interval [2π,π].
The values obtained for negative frequencies L(e2jjωTj) are simply the complex
conjugate of the values L(ejjωTj) and need not be separately calculated. Because
the order of the numerator is equal to that of the denominator or less, points on
the large circle map to zero or to a single constant value. Because the value ε is
infinitesimal, the values on the straight-line portions close to the real axis cancel.
We can simplify the test by plotting L(ejωT) as we traverse the contour and
then counting its encirclements of the point 21 1 j 0. As Figure 4.8 shows, this is
equivalent to plotting pcl(z) and counting encirclements of the origin.
4.6 Nyquist criterion
ε
FIGURE 4.7
Contour for stability determination.
0.6
Imaginary Axis
0.4
1+ L
L
0.2
0
–0.2
–0.4
–0.6
–1
–0.5
0
0.5
Real Axis
1
1.5
2
FIGURE 4.8
Nyquist plots of L and 1 1 L.
If the system has open-loop poles on the unit circle, the contour passes
through poles and the test fails. To avoid this, we modify the contour to avoid
these open-loop poles. The most common case is a pole at unity for which
the modified contour is shown in Figure 4.9. The contour includes an additional
circular arc of infinitesimal radius. Because the portions on the positive real
111
112
CHAPTER 4 Stability of Digital Control Systems
ε
FIGURE 4.9
Modified contour for stability determination.
•
FIGURE 4.10
Simplification of the modified contour for stability determination with (1, 0) shown in gray.
axis cancel, the contour effectively reduces to the one shown in Figure 4.10. For
m poles at unity, the loop gain is given by
LðzÞ 5
NL ðsÞ
ðz 21Þm DðsÞ
(4.27)
where NL and D have no unity roots. The value of the transfer function on the
circular arc is approximately given by
LðzÞz-11εejθ 5
K
;
εejθm
h π πi
θA 2 ;
2 2
(4.28)
where K is equal to NL(1)/D(1) and (z 2 1) 5 εe jq, with ε the radius of the small
circular arc. Therefore, the small circle maps to a large circle, and traversing the
4.6 Nyquist criterion
small circle once causes a net denominator angle change of 2 mπ radians (clockwise) (i.e., m half circles). The net angle change for the quotient on traversing the
small circular arc is thus mp radians (counterclockwise). We conclude that for a
type m system, the Nyquist contour will include m large clockwise semicircles.
We now summarize the results obtained in Theorem 4.7.
THEOREM 4.7: NYQUIST CRITERION
Let the number of counterclockwise encirclements of the point (21, 0) for a loop gain L(z)
when traversing the stability contour be N (i.e., 2N for clockwise encirclements), where L(z)
has P open-loop poles inside the contour. Then the system has Z closed-loop poles outside
the unit circle with Z given by
Z 5 ð2NÞ 1 P
(4.29)
COROLLARY
An open-loop stable system is closed-loop stable if and only if its Nyquist plot does not encircle
the point (21, 0) (i.e., if N 5 0).
Although counting encirclements appears complicated, it is actually quite
simple using the following recipe:
1. Starting at a distant point, move toward the point (21, 0).
2. Count all lines of the stability contour crossed. Count each line with an arrow
pointing from your left to your right as negative and every line with an arrow
pointing from your right to your left as positive.
3. The net number of lines counted is equal to the number of encirclements of
the point (21, 0).
The recipe is demonstrated in Example 4.9.
EXAMPLE 4.9
Consider a linearized model of a furnace:
GðsÞ 5
Ti ðsÞ
Kgrw giw
5 2
s 1ð2giw 1 grw Þs 1 grw giw
UðsÞ
During the heating phase, we have the model2
GðsÞ 5
1
s2 1 3s 1 1
Determine the closed-loop stability of the system with digital control and a sampling
period of 0.01.
113
CHAPTER 4 Stability of Digital Control Systems
Im[G(jω)]
0.8
0.6
0.4
Imaginary Axis
114
0.2
0
–0.2
–0.4
–0.6
–0.8
–1
–0.8 –0.6 –0.4 –0.2 0
0.2
Real Axis
0.4
0.6
0.8
1 Re[G(jω)]
FIGURE 4.11
Nyquist plot of the furnace transfer function.
Solution
In Example 3.5, we determined the transfer function of the system to be
GZAS ðzÞ 5 K
p1 e2p2 T 2 p2 e2p1 T 1 p2 2 p1 z 1 p1 e2p1 2 p2 e2p2 1 p2 e2ðp1 1p2 Þ 2 p1 e2ðp1 1p2 Þ
p1 p2 ðp2 2 p1 Þðz 2 e2p1 T Þðz 2 e2p2 T Þ
Substituting the values of the parameters in the general expression gives the z-transfer
function
GZAS ðzÞ 5 1025
4:95z 1 4:901
z2 2 1:97z 1 0:9704
The Nyquist plot of the system is shown in Figure 4.11. The plot shows that the Nyquist
plot does not encircle the point (21, 0)—that is, N 5 0. Because the open-loop transfer
function has no RHP poles (P 5 0), the system is closed-loop stable (Z 5 0). Note that there
is a free path to the point (21, 0) without crossing any lines of the plot because there are
no encirclements.
2
Hagglund, T., Tengall, A., 1995. An automatic tuning procedure for unsymmetrical processes, Proceedings of
the European Control Conference, Rome, 1995.
4.6.1 Phase margin and gain margin
In practice, the stability of a mathematical model is not sufficient to guarantee
acceptable system performance or even to guarantee the stability of the physical
4.6 Nyquist criterion
system that the model represents. This is because of the approximate nature of
mathematical models and our imperfect knowledge of the system parameters. We
therefore need to determine how far the system is from instability. This degree of
stability is known as relative stability. To keep our discussion simple, we restrict
it to open-loop stable systems where zero encirclements guarantee stability. For
open-loop stable systems that are nominally closed-loop stable, the distance from
instability can be measured by the distance between the set of points of the
Nyquist plot and the point (21, 0).
Typically, the distance between a set of points and the single point (21, 0) is
defined as the minimum distance over the set of points. However, it is more
convenient to define relative stability in terms of two distances: a magnitude
distance and an angular distance. The two distances are given in the following
definitions.
DEFINITION 4.5: GAIN MARGIN
The gain margin is the gain perturbation that makes the system marginally stable.
DEFINITION 4.6: PHASE MARGIN
The phase margin is the negative phase perturbation that makes the system marginally stable.
The two stability margins become clearer by examining the block diagram
of Figure 4.12. If the system has a multiplicative gain perturbation ∆G(z) 5 ∆K,
then the gain margin is the magnitude of ∆K that makes the system on the verge
of instability. If the system has a multiplicative gain perturbation ∆G(z) 5 e2j∆θ,
then the gain margin is the lag angle ∆θ that makes the system on the verge of
instability. Clearly, the perturbations corresponding to the gain margin and phase
margin are limiting values, and satisfactory behavior would require smaller model
perturbations.
The Nyquist plot of Figure 4.13 shows the gain margin and phase margin for a
given polar plot (the positive frequency portion of the Nyquist plot). Recall that each
point on the plot represents a complex number, which is represented by a vector
from the origin. Scaling the plot with a gain ∆K results in scaled vectors without
rotation. Thus, the vector on the negative real axis is the one that reaches the point
R(s) +
−
FIGURE 4.12
Model perturbation ∆G(s).
∆G(z)
Y(z)
G(z)
115
CHAPTER 4 Stability of Digital Control Systems
Im[G(jω)]
1
1/GM
0
PM
Imaginary Axis
116
–1
–2
–3
–4
–5
–2
–1.8 –1.6 –1.4 –1.2
–1
–0.8 –0.6 –0.4 –0.2
0
Re[G(jω)]
Real Axis
FIGURE 4.13
Nyquist plot with phase margin and gain margin.
(21, 0) if appropriately scaled, and the magnitude of that vector is the reciprocal of
the gain margin. On the other hand, multiplication by e2j∆θ rotates the plot clockwise
without changing the magnitudes of the vectors, and it is the vector of magnitude
unity that can reach the point (21, 0) if rotated by the phase margin.
For an unstable system, a counterclockwise rotation or a reduction in gain
is needed to make the system on the verge of instability. The system will have
a negative phase margin and a gain margin less than unity, which is also negative if it is expressed in decibels—that is, in units of 20 log{jG(jω)j}. The polar
plot of a system with negative gain margin and phase margin is shown in
Figure 4.14.
The gain margin can be obtained analytically by equating the imaginary part
of the frequency response to zero and solving for the real part. The phase margin
can be obtained by equating the magnitude of the frequency response to unity
and solving for angle, and then adding 180 . However, because only approximate
values are needed in practice, it is easier to use MATLAB to obtain both margins.
In some cases, the intercept with the real axis can be obtained as the value where
z 5 21 provided that the system has no pole at 21 (i.e., the frequency response
has no discontinuity at the folding frequency ωs/2).
It is sometimes convenient to plot the frequency response using the Bode
plot, but the availability of frequency response plotting commands in MATLAB
reduces the necessity for such plots. The MATLAB commands for obtaining
4.6 Nyquist criterion
Im[G(jω)]
1
Imaginary Axis
0.5
0
–0.5
–1
–1.5
–2
–3
–2.5
–2
–1.5
–1
Real Axis
–0.5
0
Re[G(jω)]
FIGURE 4.14
Nyquist plot with negative gain margin (dBs) and phase margin.
frequency response plots (which work for both continuous-time and discretetime systems) are
.. nyquistðgdÞ %Nyquist plot
.. bodeðgdÞ %Bode plot
It is also possible to find the gain and phase margins with the command
.. ½gm;pm 5 marginðgdÞ
An alternative form of the command is
.. marginðgdÞ
The latter form shows the gain margin and phase margin on the Bode plot of the
system. We can also obtain the phase margin and gain margin using the Nyquist
plot by clicking on the plot and selecting
Characteristics
All stability margins
The concepts of the gain margin and phase margin and their evaluation using
MATLAB are illustrated in Example 4.10.
117
CHAPTER 4 Stability of Digital Control Systems
EXAMPLE 4.10
Determine the closed-loop stability of the digital control system for the furnace model of
Example 4.9 with a discrete-time first-order actuator of the form
Ga ðzÞ 5
0:9516
z 2 0:9048
and a sampling period of 0.01. If an amplifier of gain K 5 5 is added to the actuator,
how does the value of the gain affect closed-loop stability?
Solution
We use MATLAB to obtain the z-transfer function of the plant and actuator:
GZAS ðzÞ 5 1025
4:711z 1 4:644
z3 2 2:875z2 1 2:753z 2 0:8781
The Nyquist plot for the system, Figure 4.15, is obtained with no additional gain and
then for a gain K 5 5. We also show the plot in the vicinity of the point (21, 0) in
Figure 4.16, from which we see that the system with K 5 5 encircles the point twice
clockwise.
Im[G(jω)]
25
20
15
10
Imaginary Axis
118
5
0
–5
–10
–15
–20
–25
–5
0
5
10
15
20 25
Real Axis
30
35
40
45
FIGURE 4.15
Nyquist plot for the furnace and actuator (K 5 1, black, K 5 5, gray).
50 Re[G(jω)]
4.6 Nyquist criterion
Im[G(jω)]
2
1.5
Imaginary Axis
1
System: untitled1
Real: –1.41
Imag: 0.000741
Frequency (rad/sec): –5.33
0.5
0
–0.5
System: gdt
Real: –0.795
Imag: –0.624
Frequency (rad/sec): 2.59
–1
–1.5
–2
–5
–4.5
–4
–3.5
–3
–2.5 –2
Real Axis
–1.5
–1
–0.5
0 Re[G(jω)]
FIGURE 4.16
Nyquist plot for the furnace and actuator in the vicinity of the point (21, 0) (K 5 1, black,
K 5 5, gray).
We count the encirclements by starting away from the point (21, 0) and counting the
lines crossed as we approach it. We cross the gray curve twice, and at each crossing
the arrow indicates that the line is moving from our right to our left (i.e., two clockwise
encirclements). The system is unstable, and the number of closed-loop poles outside the
unit circle is given by
Z 5 ð2NÞ 1 P
5210
For the original gain of unity, the intercept with the real axis is at a magnitude of
approximately 0.28 and can be increased by a factor of about 3.5 before the system
becomes unstable.
At a magnitude of unity, the phase is about 38 less negative than the
instability value of 2180 . We therefore have a gain margin of about 3.5 and a phase
margin of about 38 . Using MATLAB, we find approximately the same values for the
margins:
.. ½gm;pm 5 marginðgtdÞ
gm 5 3:4817
pm 5 37:5426
119
Magnitude (dB)
CHAPTER 4 Stability of Digital Control Systems
Phase (deg)
120
20
10
0
–10
–20
–30
–40
–50
–60
–70
–80
GM = 10.8 dB (at 6.25 rad/sec), PM = 37.5 deg (at 2.6 rad/sec)
0
–45
–90
–135
–180
–225
–270
–315
–360
–1
10
GM
PM
0
1
10
10
Frequency (rad/sec)
2
10
FIGURE 4.17
Phase margin and gain margin for the oven control system shown on the Bode plot.
Thus, an additional gain of over 3 or an additional phase lag of over 37 can be tolerated
without causing instability. However, such perturbations may cause a significant deterioration in the time response of the system. Perturbations in gain and phase may actually
occur upon implementing the control, and the margins are needed for successful implementation. In fact, the phase margin of the system is rather low, and a controller may be
needed to improve the response of the system.
To obtain the Bode plot showing the phase margin and gain margin of Figure 4.17,
we use the following command:
.. marginðgtdÞ
1. The phase margin for unity gain shown on the plot is as obtained with the first form
of the command margin, but the gain margin is in dBs. The values are nevertheless
identical as verified with the MATLAB command:
.. 20 log10ðgmÞ
ans 5 10:8359
2. The gain margin and phase margin can also be obtained using the Nyquist command
as shown in Figure 4.18.
4.6 Nyquist criterion
Im[G(jω)]
6
4
Imaginary Axis
2
System: gdt
Gain Margin (dB): 10.8
At frequency (rad/sec): 5.25
Closed-Loop Stable? Yes
0
System: gdt
Phase Margin (deg): 37.5
Delay Margin (samples): 25.2
At frequency (rad/sec): 2.6
Closed-Loop Stable? Yes
–2
–4
–6
–1
–0.9 –0.8 –0.7 –0.6 –0.5 –0.4 –0.3 –0.2 –0.1
Real Axis
0 Re[G(jω)]
FIGURE 4.18
Phase margin and gain margin for the oven control system shown on the Nyquist plot.
EXAMPLE 4.11
Determine the closed-loop stability of the digital control system for the position control
system with analog transfer function
GðsÞ 5
10
sðs 1 1Þ
and with a sampling period of 0.01. If the system is stable, determine the gain margin
and the phase margin.
Solution
We first obtain the transfer function for the analog plant with ADC and DAC. The transfer
function is given by
GZAS ðzÞ 5 4:983 3 1024
z 1 0:9967
ðz 2 1Þðz 2 0:99Þ
Note that the transfer function has a pole at unity because the analog transfer function
has a pole at the origin or is type I. Although such systems require the use of the modified
Nyquist contour, this has no significant impact on the steps required for stability testing
using the Nyquist criterion. The Nyquist plot obtained using the MATLAB command nyquist
is shown in Figure 4.19.
121
CHAPTER 4 Stability of Digital Control Systems
Im[G(jω)]
10
8
6
System: gd
Gain Margin (dB): 26
At frequency (rad/sec): 14.1
Closed-Loop Stable? Yes
4
Imaginary Axis
2
0
System: gd
Phase Margin (deg): 17.1
Delay Margin (samples): 9.67
At frequency (rad/sec): 3.08
Closed-Loop Stable? Yes
–2
–4
–6
–8
–10
–7
–6
–5
–4
–3
Real Axis
–2
–1
0 Re[G(jω)]
FIGURE 4.19
Nyquist plot for the position control system of Example 4.11.
Magnitude (dB)
100
GM = 26 dB (at 14.1 rad/sec), PM = 17.1 deg (at 3.08 rad/sec)
50
0
–50
–100
–150
–90
Phase (deg)
122
–135
–180
–225
–270
10–2
10–1
100
101
Frequency (rad/sec)
102
103
FIGURE 4.20
Bode diagram with phase margin and gain margin for the position control system
of Example 4.11.
Problems
The plot does not include the large semicircle corresponding to the small semicircle
on the modified contour in Figure 4.10. However, this does not prevent us from investigating stability. It is obvious that the contour does not encircle the point (21, 0) because the
point is to the left of the observer moving along the polar plot (lower half). In addition, we
can reach the (21, 0) point without crossing any of the lines of the Nyquist plot. The system is stable because the number of closed-loop poles outside the unit circle is given by
Z 5 ð2NÞ 1 P
501050
The gain margin is 17.1 , and the phase margin is 26 dB. The gain margin and phase
margin can also be obtained using the margin command as shown in Figure 4.20.
Resources
Franklin, G.F., Powell, J.D., Workman, M.L., 1990. Digital Control of Dynamic Systems.
Addison-Wesley, Reading, MA.
Gupta, S.C., Hasdorff, L., 1970. Fundamentals of Automatic Control. Wiley, New York.
Jury, E.I., 1973. Theory and Applications of the z-Transform Method. Krieger, Huntington, NY.
Kuo, B.C., 1992. Digital Control Systems. Saunders, Fort Worth, TX.
Ogata, K., 1987. Digital Control Engineering. Prentice Hall, Englewood Cliffs, NJ.
Oppenheim, A.V., Willsky, A.S., Young, I.T., 1983. Signals and Systems. Prentice Hall,
Englewood Cliffs, NJ.
Ragazzini, J.R., Franklin, G.F., 1958. Sampled-Data Control Systems. McGraw-Hill,
New York.
PROBLEMS
3
4.1
Determine the asymptotic stability and the BIBO stability of the following
systems:
(a) yðk 12Þ 10:8yðk 1 1Þ1 0:07yðkÞ5 2uðk 1 1Þ 1 0:2uðkÞ k 5 0; 1; 2;.. .
(b) yðk 12Þ 20:8yðk 1 1Þ1 0:07yðkÞ 5 2uðk 1 1Þ 1 0:2uðkÞ k 5 0; 1; 2;...
(c) yðk 1 2Þ 1 0:1yðk 1 1Þ 1 0:9yðkÞ 5 3:0uðkÞ
k 5 0; 1; 2; . . .
4.2
Biochemical reactors are used in different processes such as waste
treatment and alcohol fermentation. By considering the dilution rate as the
manipulated variable and the biomass concentration as the measured
output, the biochemical reactor can be modeled by the following transfer
function in the vicinity of an unstable steady-state operating point:3
5:8644
GðsÞ 5
25:888s 1 1
Bequette, B.W., 2003. Process Control: Modeling, Design, and Simulation, Prentice Hall, Upper
Saddle River, NJ.
123
124
CHAPTER 4 Stability of Digital Control Systems
Determine GZAS(z) with a sampling rate T 5 0.1, and then consider the
feedback controller
CðzÞ 5 2
z 2 1:017
z21
Verify that the resulting feedback system is not internally stable.
4.3
Use the Routh-Hurwitz criterion to investigate the stability of the following
systems:
5ðz 2 2Þ
(a) GðzÞ 5
ðz 2 0:1Þðz 2 0:8Þ
(b) GðzÞ 5
10ðz 1 0:1Þ
ðz 2 0:7Þðz 2 0:9Þ
4.4
Repeat Problem 4.3 using the Jury criterion.
4.5
Obtain the impulse response for the systems shown in Problem 4.3, and
verify the results obtained using the Routh-Hurwitz criterion. Also determine
the exponential rate of decay for each impulse response sequence.
4.6
Use the Routh-Hurwitz criterion to find the stable range of K for the
closed-loop unity feedback systems with loop gain
Kðz 2 1Þ
(a) GðzÞ 5
ðz 2 0:1Þðz 2 0:8Þ
(b) GðzÞ 5
Kðz 1 0:1Þ
ðz 2 0:7Þðz 2 0:9Þ
4.7
Repeat Problem 4.6 using the Jury criterion.
4.8
Use the Jury criterion to determine the stability of the following
polynomials:
(a) z5 1 0.2z4 1 z2 1 0.3z 2 0.1 5 0
(b) z5 2 0.25z4 1 0.1z3 1 0.4z2 1 0.3z 2 0.1 5 0
4.9
Determine the stable range of the parameter a for the closed-loop unity
feedback systems with loop gain
1:1ðz 2 1Þ
(a) GðzÞ 5
ðz 2 aÞðz 2 0:8Þ
(b) GðzÞ 5
4.10
1:2ðz 1 0:1Þ
ðz 2 aÞðz 2 0:9Þ
For a gain of 0.5, derive the gain margin and phase margin of the systems of Problem 4.6 analytically. Let T 5 1 with no loss of generality
because the value of ωT in radians is all that is needed for the solution.
Explain why the phase margin is not defined for the system shown in
Computer Exercises
Problem 4.6(a). Hint: The gain margin is obtained by finding the point
where the imaginary part of the frequency response is zero. The phase
margin is obtained by finding the point where the magnitude of the
frequency response is unity.
COMPUTER EXERCISES
4.11
Write a computer program to perform the Routh-Hurwitz test using a
suitable CAD tool.
4.12
Write a computer program to perform the Jury test using a suitable CAD tool.
4.13
Write a computer program that uses the Jury test program in Exercise 4.12
to determine the stability of a system with an uncertain gain K in a given
range [Kmin, Kmax]. Verify the answers obtained for Problem 4.6 using
your program.
4.14
Show how the program written for Exercise 4.13 can be used to test the
stability of a system with uncertain zero location. Use the program to test
the effect of a 620% variation in the location of the zero for the systems
shown in Problem 4.6, with a fixed gain equal to half the critical value.
4.15
Show how the program written for Exercise 4.13 can be used to test the
stability of a system with uncertain pole location. Use the program to
test the effect of a 620% variation in the location of the first pole for the
systems shown in Problem 4.6, with a fixed gain equal to half the critical
value.
4.16
Simulate the closed-loop systems shown in Problem 4.6 with a unit step
input and (a) gain K equal to half the critical gain and (b) gain K equal to
the critical gain. Discuss their stability using your simulation results.
4.17
For unity gain, obtain the Nyquist plots of the systems shown in Problem
4.6 using MATLAB and determine the following:
(a) The intersection with the real axis using the Nyquist plot and then
using the Bode plot
(b) The stable range of positive gains K for the closed-loop unity feedback
systems
(c) The gain margin and phase margin for a gain K 5 0.5
4.18
For twice the nominal gain, use MATLAB to obtain the Nyquist and Bode
plots of the systems of the furnace control system of Example 4.10 with a
sampling period of 0.01 and determine the following:
(a) The intersection with the real axis using the Nyquist plot and then
using the Bode plot
125
126
CHAPTER 4 Stability of Digital Control Systems
(b) The stable range of additional positive gains K for the closed-loop
unity feedback systems
(c) The gain margin and phase margin for twice the nominal gain
4.19
In many applications, there is a need for accurate position control at the
nanometer scale. This is known as nanopositioning and is now feasible
because of advances in nanotechnology. The following transfer function
represents a single-axis nanopositioning system:4
GðsÞ 5
4:29 3 1010 ðs2 1 631:2s 1 9:4 3 106 Þ
ðs2 1 178:2s 1 6 3 106 Þðs2 1 412:3s 1 16 3 106 Þ
ðs2
ðs2 1 638:8s 1 45 3 106 Þ
1 209:7s 1 56 3 106 Þðs 1 5818Þ
(a) Obtain the DAC-analog system-ADC transfer function for a sampling
period of 100 ms, and determine its stability using the Nyquist criterion.
(b) Obtain the DAC-analog system-ADC transfer function for a sampling
period of 1 ms, and determine its stability using the Nyquist criterion.
(c) Plot the closed-loop step response of the system of (b), and explain the
stability results of (a) and (b) based on your plot.
4
Sebastian, A., Salapaka, S.M., 2005. Design methodologies of robust nano-positioning, IEEE
Trans. Control Systems Tech. 13 (6), pp. 868876.
CHAPTER
Analog Control System
Design
5
OBJECTIVES
After completing this chapter, the reader will be able to do the following:
1. Obtain root locus plots for analog systems.
2. Characterize a system’s step response based on its root locus plot.
3. Design proportional (P), proportional-derivative (PD), proportional-integral (PI),
and proportional-integral-derivative (PID) controllers in the s-domain.
4. Tune PID controllers using the Ziegler-Nichols approach.
Analog controllers can be implemented using analog components or approximated
with digital controllers using standard analog-to-digital transformations. In addition,
direct digital control system design in the z-domain is very similar to the s-domain
design of analog systems. Thus, a review of classical control design is the first step
toward understanding the design of digital control systems. This chapter reviews
the design of analog controllers in the s-domain and prepares the reader for the
digital controller design methods presented in Chapter 6. The reader is assumed to
have had some exposure to the s-domain and its use in control system design.
5.1 Root locus
The root locus method provides a quick means of predicting the closed-loop
behavior of a system based on its open-loop poles and zeros. The method is based
on the properties of the closed-loop characteristic equation
1 1 KLðsÞ 5 0
(5.1)
where the gain K is a design parameter and L(s) is the loop gain of the system.
We assume a loop gain of the form
nz
L ðs 2 zi Þ
LðsÞ 5
i51
np
(5.2)
L ðs 2 pj Þ
j51
127
128
CHAPTER 5 Analog Control System Design
where zi, i 5 1, 2,. . ., nz, are the open-loop system zeros and pj, j 5 1, 2,. . ., np are
the open-loop system poles. It is required to determine the loci of the closed-loop
poles of the system (root loci) as K varies between zero and infinity.1 Because
of the relationship between pole locations and the time response, this gives a
preview of the closed-loop system behavior for different K.
The complex equality (5.1) is equivalent to the two real equalities:
•
•
Magnitude condition KjL(s)j 5 1
Angle condition +L(s) 5 6(2m 1 1)180 ,
m 5 0, 1, 2, . . .
Using (5.1) or the preceding conditions, the following rules for sketching root
loci can be derived:
1. The number of root locus branches is equal to the number of open-loop poles
of L(s).
2. The root locus branches start at the open-loop poles and end at the open-loop
zeros or at infinity.
3. The real axis root loci have an odd number of poles plus zeros to their right.
4. The branches going to infinity asymptotically approach the straight lines
defined by the angle
θa 5 6
ð2m 1 1Þ180
;
np 2 nz
m 5 0; 1; 2; . . .
(5.3)
and the intercept
np
P
σa 5
i51
pi 2
nz
P
zj
j51
(5.4)
np 2 nz
5. Breakaway points (points of departure from the real axis) correspond to local
maxima of K, whereas break-in points (points of arrival at the real axis)
correspond to local minima of K.
6. The angle of departure from a complex pole pn is given by
180 2
nX
p 21
i51
+ðpn 2 pi Þ 1
nz
X
+ðpn 2 zj Þ
(5.5)
j51
The angle of arrival at a complex zero is similarly defined.
1
In rare cases, negative gain values are allowed, and the corresponding root loci are obtained.
Negative root loci are not addressed in this text.
5.1 Root locus
EXAMPLE 5.1
Sketch the root locus plots for the loop gains
1
ðs 1 1Þðs 1 3Þ
1
2. LðsÞ 5
ðs 1 1Þðs 1 3Þðs 1 5Þ
s15
3. LðsÞ 5
ðs 1 1Þðs 1 3Þ
1. LðsÞ 5
Comment on the effect of adding a pole or a zero to the loop gain.
Solution
5
4
Imaginary Axis
1
0.8
0.6
0.4
0.2
0
–0.2
–0.4
–0.6
–0.8
–1
–3
3
2
1
0
–2.5
–2
–1.5 –1
Real Axis
(a)
Imaginary Axis
Imaginary Axis
The root loci for the three loop gains as obtained using MATLAB are shown in
Figure 5.1. We now discuss how these plots can be sketched using root locus sketching
rules.
–0.5
0
–1
–2.5
–2
–1.5 –1 –0.5
Real Axis
(b)
0
4
3
2
1
0
–1
–2
–3
–4
–9 –8 –7 –6 –5 –4 –3 –2 –1 0
Real Axis
(c)
1
FIGURE 5.1
Root loci of second- and third-order systems. (a) Root locus of a second-order
system. (b) Root locus of a third-order system. (c) Root locus of a second-order
system with zero.
0.5
129
130
CHAPTER 5 Analog Control System Design
1. Using rule 1, the function has two root locus branches. By rule 2, the branches start at 21
and 23 and go to infinity. By rule 3, the real axis locus is between (21) and (23). Rule 4
gives the asymptote angles
θa 5 6
ð2m 1 1Þ180
;
2
m 5 0; 1; 2; . . .
5 6 90 ; 6 270 ; . . .
and the intercept
21 2 3
5 22
2
To find the breakaway point using Rule 5, we express real K using the characteristic
equation as
σa 5
K 5 2ðσ 1 1Þðσ 1 3Þ 5 2ðσ2 1 4σ 1 3Þ
We then differentiate with respect to σ and equate to zero for a maximum to obtain
2
dK
5 2σ 1 4 5 0
dσ
Hence, the breakaway point is at σb 5 22. This corresponds to a maximum of K
because the second derivative is equal to 22 (negative). It can be easily shown that
for any system with only two real axis poles, the breakaway point is midway between
the two poles.
2. The root locus has three branches, with each branch starting at one of the open-loop
poles (21, 23, 25). The real axis loci are between 21 and 23 and to the left of 25. The
branches all go to infinity, with one branch remaining on the negative real axis and the
other two breaking away. The breakaway point is given by the maximum of the real gain K
K 5 2ðσ 1 1Þðσ 1 3Þðσ 1 5Þ
Differentiating gives
2
dK
5 ðσ 1 1Þðσ 1 3Þ 1 ðσ 1 3Þðσ 1 5Þ 1 ðσ 1 1Þðσ 1 5Þ
dσ
5 3σ2 1 18σ 1 23
50
which yields σb 5 21.845 or 24.155. The first value is the desired breakaway point
because it lies on the real axis locus between the poles and 21 and 23. The second
value corresponds to a negative gain value and is therefore inadmissible. The gain at
the breakaway point can be evaluated from the magnitude condition and is given by
K 5 2ð21:845 1 1Þð21:845 1 3Þð21:845 1 5Þ 5 3:079
The asymptotes are defined by the angles
θa 5 6
ð2m 1 1Þ180
;
3
m 5 0; 1; 2; . . .
5 660 ; 6180 ; . . .
and the intercept by
σa 5
21 2 3 2 5
5 23
3
5.1 Root locus
The closed-loop characteristic equation
s3 1 9s2 1 25s 1 15 1 K 5 0
corresponds to the Routh table
1
s3 9
2
s 192 2 K
1
s 9
s0 15 1 K
23
15 1 K
Thus, at K 5 192, a zero row results. This value defines the auxiliary equation
9s2 1 207 5 0
Thus, the intersection with the jω-axis is 6 j4.796 rad/s. The intersection can also be
obtained by factorizing the characteristic polynomial at the critical gain
s2 1 9s2 1 25s 1 15 1 192 5 ðs 1 9Þðs 1 j4:796Þðs 2 j4:796Þ
3. The root locus has two branches as in (1), but now one of the branches ends at the
zero. From the characteristic equation, the gain is given by
K52
ðσ 1 1Þðσ 1 3Þ
σ15
Differentiating gives
dK
ðσ 1 1 1 σ 1 3Þðσ 1 5Þ 2 ðσ 1 1Þðσ 1 3Þ
52
dσ
ðσ15Þ2
52
σ2 1 10σ 1 17
ðσ15Þ2
50
which yields σb 5 22.172 or 27.828. The first value is the breakaway point because
it lies between the poles, whereas the second value is to the left of the zero and corresponds to the break-in point. The second derivative
d2 K
ð2σ 1 10Þðσ 1 5Þ 2 2ðσ2 1 10σ 1 17Þ
52
2
dσ
ðσ15Þ3
5 216=ðσ15Þ3
is negative for the first value and positive for the second value. Hence, K has a maximum at the first value and a minimum at the second. It can be shown that the root
locus is a circle centered at the zero with radius given by the geometric mean of the
distances between the zero and the two real poles.
Clearly, adding a pole pushes the root locus branches toward the RHP,
whereas adding a zero pulls them back into the LHP. Thus, adding a zero allows
the use of higher gain values without destabilizing the system. In practice, the
allowable increase in gain is limited by the cost of the associated increase in control effort and by the possibility of driving the system outside the linear range of
operation.
131
132
CHAPTER 5 Analog Control System Design
5.2 Root locus using MATLAB
While the above rules together with (5.1) allow the sketching of root loci for any
loop gain of the form (5.2), it is often sufficient to use a subset of these rules
to obtain the root loci. For higher-order or more complex situations, it is easier to
use a CAD tool like MATLAB. These packages do not actually use root locus
sketching rules. Instead, they numerically solve for the roots of the characteristic
equation as K is varied in a given range and then display the root loci.
The MATLAB command to obtain root locus plots is “rlocus.” To obtain the
root locus of the system
GðsÞ 5
s2
s15
1 2s 1 10
using MATLAB enter
.. g 5 tfð½1; 5; ½1; 2; 10Þ;
.. rlocusðgÞ;
To obtain specific points on the root locus and the corresponding data, we
simply click or the root locus. Dragging the mouse allows us to change the referenced point to obtain more data.
5.3 Design specifications and the effect of gain variation
The objective of control system design is to construct a system that has a desirable response to standard inputs. A desirable transient response is one that is
sufficiently fast without excessive oscillations. A desirable steady-state response
is one that follows the desired output with sufficient accuracy. In terms of
the response to a unit step input, the transient response is characterized by the
following criteria:
1. Time constant τ. Time required to reach about 63% of the final value.
2. Rise time Tr. Time to go from 10% to 90% of the final value.
3. Percentage overshoot (PO).
PO 5
Peak value 2 Final value
3 100%
Final value
4. Peak time Tp. Time to first peak of an oscillatory response.
5. Settling time Ts. Time after which the oscillatory response remains within a
specified percentage (usually 2 percent) of the final value.
Clearly, the percentage overshoot and the peak time are intended for use with
an oscillatory response (i.e., for a system with at least one pair of complex conjugate poles). For a single complex conjugate pair, these criteria can be expressed
in terms of the pole locations.
5.3 Design specifications and the effect of gain variation
Consider the second-order system
GðsÞ 5
s2
ω2n
1 2ζωn s 1 ω2n
(5.6)
where ζ is the damping ratio and ωn is the undamped natural frequency. Then
criteria 3 through 5 are given by
ffiffiffiffiffi2ffi
2pπζ
PO 5 e
Tp 5
12ζ
3 100%
(5.7)
π
π
5 pffiffiffiffiffiffiffiffiffiffiffiffiffi
ωd
ωn 1 2 ζ 2
(5.8)
4
ζωn
(5.9)
Ts From (5.7) and (5.9), the damping ratio ζ is an indicator of the oscillatory
nature of the response, with excessive oscillations occurring at low ζ values.
Hence, ζ is used as a measure of the relative stability of the system. From (5.8),
the time to first peak drops as the undamped natural frequency ωn increases.
Hence, ωn is used as a measure of speed of response. For higher-order systems,
these measures and equations (5.7) through (5.9) can provide approximate
answers if the time response is dominated by a single pair of complex conjugate
poles. This occurs if additional poles and zeros are far in the left half plane or
almost cancel. For systems with zeros, the percentage overshoot is higher than
predicted by (5.7) unless the zero is located far in the LHP or almost cancels with
a pole. However, equations (5.7) through (5.9) are always used in design because
of their simplicity.
Thus, the design process reduces to the selection of pole locations and the
corresponding behavior in the time domain. The root locus summarizes information on the time response of a closed-loop system as dictated by the pole locations
in a single plot. Together with the previously stated criteria, it provides a powerful
design tool, as demonstrated by the next example.
EXAMPLE 5.2
Discuss the effect of gain variation on the time response of the position control system
described in Example 3.3 with
LðsÞ 5
1
sðs 1 pÞ
Solution
The root locus of the system is similar to that of Example 5.1(1), and it is shown in
Figure 5.2(a) for p 5 4. As the gain K is increased, the closed-loop system poles become
complex conjugate; then the damping ratio ζ decreases progressively. Thus, the relative
stability of the system deteriorates for high gain values. However, large gain values are
133
134
CHAPTER 5 Analog Control System Design
Amplitude
jω
4
1
3
2
0.8
1
0.6
0
0.4
–1
–2
0.2
–3
–4
0
–4
–2
(a)
0
2 σ
0
0.5
1
1.5
2
2.5
3 Time s
(b)
FIGURE 5.2
Use of the root locus in the design of a second-order system. (a) Root locus for p 5 4.
(b) Step response for K 5 10.
required to reduce the steady-state error of the system due to a unit ramp, which is
given by
eðNÞ% 5
100 100p
5
Kv
K
In addition, increasing K increases the undamped natural frequency ωn (i.e., the magnitude of the pole), and hence the speed of response of the system increases. Thus, the
chosen gain must be a compromise value that is large enough for a low steady-state error
and an acceptable speed of response but small enough to avoid excessive oscillations. The
time response for a gain of 10 is shown in Figure 5.2(b).
Example 5.2 illustrates an important feature of design—namely, that it
typically involves a compromise between conflicting requirements. The designer
must always remember this when selecting design specifications so as to avoid
over-optimizing some design criteria at the expense of others.
Note that the settling time of the system does not change in this case when
the gain is increased. For a higher-order system or a system with a zero, this
is usually not the case. Yet, for simplicity, the second-order equations (5.7)
through (5.9) are still used in design. The designer must always be alert to errors
that this simplification may cause. In practice, design is an iterative process where
the approximate results from the second-order approximation are checked and,
if necessary, the design is repeated until satisfactory results are obtained.
5.4 Root locus design
5.4 Root locus design
Laplace transformation of a time function yields a function of the complex
variable s that contains information about the transformed time function. We can
therefore use the poles of the s-domain function to characterize the behavior of
the time function without inverse transformation. Figure 5.3 shows pole locations
in the s-domain and the associated time functions. Real poles are associated with
3
2.5
2
1.5
1
0.5
0
–0.5
–1
–1.5
0
1.4
1.2
1
0.8
0.6
0.4
0.2
0
–0.2
–0.4
5
0
5
2
1.5
1
0.5
0
–0.5
–1
–1.5
–2
0
150
100
50
0
–50
–100
–150
–200
–250
5
0
5
ω
¥
¥
¥
¥
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0
0.5
1
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
¥
¥
0
5
2
1.8
1.6
1.4
1.2
1
0.8
0.6
0.4
0.2
0
¥
σ
¥
¥
150
100
50
0
FIGURE 5.3
Pole locations and the associated time responses.
5
0
0
5
450
400
350
300
250
200
150
100
50
0
0
1
2
135
136
CHAPTER 5 Analog Control System Design
an exponential time response that decays for LHP poles and increases for RHP
poles. The magnitude of the pole determines the rate of exponential change.
A pole at the origin is associated with a unit step. Complex conjugate poles are
associated with an oscillatory response that decays exponentially for LHP poles
and increases exponentially for RHP poles. The real part of the pole determines
the rate of exponential change, and the imaginary part determines the frequency
of oscillations. Imaginary axis poles are associated with sustained oscillations.
The objective of control system design in the s-domain is to indirectly select
a desirable time response for the system through the selection of closed-loop pole
locations. The simplest means of shifting the system poles is through the use of
an amplifier or proportional controller. If this fails, then the pole locations can be
more drastically altered by adding a dynamic controller with its own open-loop
poles and zeros.
As Examples 5.1 and 5.2 illustrate, adding a zero to the system allows the
improvement of its time response because it pulls the root locus into the LHP.
Adding a pole at the origin increases the type number of the system and reduces
its steady-state error but may adversely affect the transient response. If an
improvement of both transient and steady-state performance is required, then it
Reference
Input +
Output
Controller
Plant
−
(a)
Reference
Input
+
Output
Plant
−
Controller
(b)
Reference
Input +
−
Cascade
Controller
+
Actuator
and Plant
Output
+
Feedback
Controller
(c)
FIGURE 5.4
Control configurations. (a) Cascade compensation. (b) Feedback compensation.
(c) Inner-loop feedback compensation.
5.4 Root locus design
may be necessary to add two zeros as well as a pole at the origin. At times, more
complex controllers may be needed to achieve the desired design objectives.
The controller could be added in the forward path, in the feedback path, or
in an inner loop. A prefilter could also be added before the control loop to allow
more freedom in design. Several controllers could be used simultaneously, if necessary, to meet all the design specifications. Examples of these control configurations
are shown in Figure 5.4.
In this section, we review the design of analog controllers. We restrict the
discussion to proportional (P), proportional-derivative (PD), proportional-integral
(PI), and proportional-integral-derivative (PID) control. Similar procedures can be
developed for the design of lead, lag, and lag-lead controllers.
5.4.1 Proportional control
Gain adjustment or proportional control allows the selection of closed-loop pole
locations from among the poles given by the root locus plot of the system loop
gain. For lower-order systems, it is possible to design proportional control systems
analytically, but a sketch of the root locus is still helpful in the design process, as
seen from Example 5.3.
EXAMPLE 5.3
A position control system (see Example 3.3) with load angular position as output and
motor armature voltage as input consists of an armature-controlled DC motor driven by a
power amplifier together with a gear train. The overall transfer function of the system is
GðsÞ 5
K
sðs 1 pÞ
Design a proportional controller for the system to obtain
1. A specified damping ratio ζ
2. A specified undamped natural frequency ωn
Solution
The root locus of the system was discussed in Example 5.2 and shown in Figure 5.2(a).
The root locus remains in the LHP for all positive gain values. The closed-loop characteristic
equation of the system is given by
sðs 1 pÞ 1 K 5 s2 1 2ζωn s 1 ω2n 5 0
Equating coefficients gives
p 5 2ζωn
K 5 ω2n
which can be solved to yield
ωn 5
pffiffiffiffi
p
K ζ 5 pffiffiffiffi
2 K
Clearly, with one free parameter either ζ or ωn can be selected, but not both. We now
select a gain value that satisfies the design specifications.
137
138
CHAPTER 5 Analog Control System Design
1. If ζ is given and p is known, then the gain of the system and its undamped natural
frequency are obtained from the equations
2
p
p
ωn 5
K5
2ζ
2ζ
2. If ωn is given and p is known, then the gain of the system and its damping ratio are
obtained from the equations
K 5 ω2n
ζ5
p
2ωn
Example 5.3 reveals some of the advantages and disadvantages of proportional
control. The design is simple, and this simplicity carries over to higher-order systems
if a CAD tool is used to assist in selecting the pole locations. Using cursor commands, the CAD tools allow the designer to select desirable pole locations from the
root locus plot directly. The step response of the system can then be examined using
the MATLAB command step. But the single free parameter available limits the
designer’s choice to one design criterion. If more than one aspect of the system time
response must be improved, a dynamic controller is needed.
5.4.2 PD control
As seen from Example 5.1, adding a zero to the loop gain improves the time
response in the system. Adding a zero is accomplished using a cascade or feedback controller of the form
CðsÞ 5 Kp 1 Kd s 5 Kd ðs 1 aÞ
a 5 Kp =Kd
(5.10)
This is known as a proportional-derivative, or PD, controller. The derivative term is only approximately realizable and is also undesirable because differentiating a noisy input results in large errors. However, if the derivative of the
output is measured, an equivalent controller is obtained without differentiation.
Thus, PD compensation is often feasible in practice.
The design of PD controllers depends on the specifications given for the
closed-loop system and on whether a feedback or cascade controller is used. For
a cascade controller, the system block diagram is shown in Figure 5.4(a) and the
closed-loop transfer function is of the form
Gcl ðsÞ 5
5
GðsÞCðsÞ
1 1 GðsÞCðsÞ
Kd ðs 1 aÞNðsÞ
DðsÞ 1 Kd ðs 1 aÞNðsÞ
(5.11)
where N(s) and D(s) are the numerator and denominator of the open-loop gain,
respectively. Pole-zero cancellation occurs if the loop gain has a pole at (2a). In
5.4 Root locus design
R(s)
Preamplifier +
Kp
−
Amplifier
Ka
Plant
G(s)
Y(s)
PD Compensator
Kd(s + a)
FIGURE 5.5
Block diagram of a PD-feedback-compensated system.
the absence of pole-zero cancellation, the closed-loop system has a zero at (2a),
which may drastically alter the time response of the system. In general, the zero
results in greater percentage overshoot than is predicted using (5.7).
Figure 5.5 shows feedback compensation including a preamplifier in cascade
with the feedback loop and an amplifier in the forward path. We show that both
amplifiers are often needed. The closed-loop transfer function is
Gcl ðsÞ 5
5
Kp Ka GðsÞ
1 1 Ka GðsÞCðsÞ
Kp Ka NðsÞ
DðsÞ 1 Ka Kd ðs 1 aÞNðsÞ
(5.12)
where Ka is the feedforward amplifier gain and Kp is the preamplifier gain. Note
that although the loop gain is the same for both cascade and feedback compensation, the closed-loop system does not have a zero at (2a) in the feedback case.
If D(s) has a pole at (2a), the feedback-compensated system has a closed-loop
pole at (2a) that appears to cancel with a zero in the root locus when in reality
it does not.
In both feedback and cascade compensation, two free design parameters are
available and two design criteria can be selected. For example, if the settling time
and percentage overshoot are specified, the steady-state error can only be checked
after the design is completed and cannot be independently chosen.
EXAMPLE 5.4
Design a PD controller for the type 1 system described in Example 5.3 to meet the following
specifications:
1. Specified ζ and ωn
2. Specified ζ and steady-state error e(N)% due to a ramp input
Consider both cascade and feedback compensation and compare them using a numerical
example.
139
140
CHAPTER 5 Analog Control System Design
Solution
The root locus of the PD-compensated system is of the form of Figure 5.1(c). This shows
that the system gain can be increased with no fear of instability. Even though this example
is solved analytically, a root locus plot is needed to give the designer a feel for the variation
in pole locations with gain.
With a PD controller the closed-loop characteristic equation is of the form
s2 1 ps 1 Kðs 1 aÞ 5 s2 1ðp 1 KÞs 1 Ka
5 s2 1 2ζωn s 1 ω2n
where K 5 Kd for cascade compensation and K 5 Ka Kd for feedback compensation.
Equating coefficients gives the equations
Ka 5 ω2n
p 1 K 5 2ζωn
1. In this case, there is no difference between (K, a) in cascade and in feedback compensation. But the feedback case requires a preamplifier with the correct gain to yield
zero steady-state error due to unit step. We examine the steady-state error in part 2.
In either case, solving for K and a gives
K 5 2ζωn 2 p a 5
ω2n
2ζωn 2 p
2. For cascade compensation, the velocity error constant of the system is
Kv 5
Ka
100
5
p
eðNÞ%
The undamped natural frequency is fixed at
pffiffiffiffiffiffi pffiffiffiffiffiffiffiffi
ωn 5 Ka 5 pKv
Solving for K and a gives
K 5 2ζ
pffiffiffiffiffiffiffiffi
pKv 2 p a 5
pKv
pffiffiffiffiffiffiffiffi
2ζ pKv 2 p
For feedback compensation with preamplifier gain Kp and cascade amplifier gain Ka,
as in Figure 5.5, the error is given by
Kp Ka
RðsÞ 2 YðsÞ 5 RðsÞ 1 2 2
s 1 ðp 1 KÞs 1 Ka
5 RðsÞ
s2 1 ðp 1 KÞs 1 Ka 2 Kp Ka
s2 1 ðp 1 KÞs 1 Ka
Using the final value theorem gives the steady-state error due to a unit ramp input as
eðNÞ% 5 L im s
s-0
1 s2 1ðp 1 KÞs 1 Ka 2 Kp Ka
3 100%
s2 1ðp 1 KÞs 1 Ka
s2
This error is infinite unless the amplifier gain is selected such that KpKa 5 Ka. The steadystate error is then given by
eðNÞ% 5
p1K
3 100%
Ka
5.4 Root locus design
The steady-state error e(N) is simply the percentage error divided by 100. Hence, using
the equations governing the closed-loop characteristic equation
Ka 5
p1K
2ζωn
5 ω2n
5
eðNÞ
eðNÞ
the undamped natural frequency is fixed at
ωn 5
2ζ
eðNÞ
Then solving for K and a we obtain
K5
4ζ 2
2p
eðNÞ
a5
4ζ 2
eðNÞð4ζ 2 2 peðNÞÞ
Note that, unlike cascade compensation, ωn can be freely selected if the steady-state
error is specified and ζ is free. To further compare cascade and feedback compensation,
we consider the system with the pole p 5 4, and require ζ 5 0.7 and ωn 5 10 rad/s for part 1.
These values give K 5 Kd 5 10 and a 5 10. In cascade compensation, (5.11) gives the
closed-loop transfer function
Gcl ðsÞ 5
10ðs 1 10Þ
s2 1 14s 1 100
For feedback compensation, amplifier gains are selected such that the numerator is
equal to 100 for unity steady-state output due to unit step input. For example, one may
select
Kp 5 10; Ka 5 10
Kd 5 1; a 5 10
Substituting in (5.12) gives the closed-loop transfer function
Gcl ðsÞ 5
100
s2 1 14s 1 100
The responses of the cascade- and feedback-compensated systems are shown together in
Figure 5.6. The PO for the feedback case can be predicted exactly using (5.7) and is equal
to about 4.6%. For cascade compensation, the PO is higher due to the presence of the
zero. The zero is at a distance from the imaginary axis less than one and a half times
the negative real part of the complex conjugate poles. Therefore, its effect is significant,
and the PO increases to over 10% with a faster response.
For part 2 with p 5 4, we specify ζ 5 0.7 and a steady-state error of 4%. Cascade
compensation requires K 5 10, a 5 10. These are identical to the values of part 1 and correspond to an undamped natural frequency ωn 5 10 rad/s.
For feedback compensation we obtain K 5 45, a 5 27.222. Using (5.12) gives the
closed-loop transfer function
Gcl ðsÞ 5
1225
s2 1 49s 1 1225
with ωn 5 35 rad/s.
The responses of cascade- and feedback-compensated systems are shown in Figure 5.7.
The PO for the feedback-compensated case is still 4.6% as calculated using (5.7). For
141
CHAPTER 5 Analog Control System Design
1.2
1
Output
0.8
0.6
0.4
0.2
0
0
0.2
0.4
0.6
0.8
1
Time s
FIGURE 5.6
Step response of PD cascade (dotted) and feedback (solid) compensated systems with a
given damping ratio and undamped natural frequency.
1.2
1
0.8
Output
142
0.6
0.4
0.2
0
0
0.2
0.4
0.6
0.8
1
Time s
FIGURE 5.7
Step response of PD cascade (dotted) and feedback (solid) compensated systems with a
given damping ratio and steady-state error.
5.4 Root locus design
cascade compensation, the PO is higher due to the presence of the zero that is close to the
complex conjugate poles.
Although the response of the feedback-compensated system is superior, several
amplifiers are required for its implementation with high gains. The high gains may cause
nonlinear behavior such as saturation in some situations.
Having demonstrated the differences between cascade and feedback compensation, we restrict our discussion in the sequel to cascade compensation. Similar
procedures can be developed for feedback compensation.
Example 5.4 is easily solved analytically because the plant is only second
order. For higher-order systems, the design is more complex, and a solution using
CAD tools is preferable. We develop design procedures using CAD tools based
on the classical graphical solution methods. These procedures combine the convenience of CAD tools and the insights that have made graphical tools successful.
The procedures find a controller transfer function such that the angle of its product with the loop gain function at the desired pole location is an odd multiple
of 180 . From the angle condition, this ensures that the desired location is on the
root locus of the compensated system. The angle contribution required from the
controller for a desired closed-loop pole location scl is
θC 5 6180 2 +Lðscl Þ
(5.13)
where L(s) is the open-loop gain with numerator N(s) and denominator D(s). For
a PD controller, the controller angle is simply the angle of the zero at the desired
pole location. Applying the angle condition at the desired closed-loop location,
it can be shown that the zero location is given by
a5
ωd
1 ζωn
tanðθC Þ
(5.14)
The proof of (5.13) and (5.14) is straightforward and is left as an exercise (see
Problem 5.5).
In some special cases, a satisfactory design is obtained by cancellation, or
near cancellation, of a system pole with the controller zero. The desired specifications are then satisfied by tuning the reduced transfer function’s gain. In practice,
exact pole-zero cancellation is impossible. However, with near cancellation the
effect of the pole-zero pair on the time response is usually negligible.
A key to the use of powerful calculators and CAD tools in place of graphical
methods is the ease with which transfer functions can be evaluated for any
complex argument using direct calculation or cursor commands. The following
CAD procedure exploits the MATLAB command evalfr to obtain the design
parameters for a specified damping ratio and undamped natural frequency. The
command evalfr evaluates a transfer function g for any complex argument s as
follows:
.. evalfrðg; sÞ
143
144
CHAPTER 5 Analog Control System Design
The angle command gives the angle of any complex number. The complex
value and its angle can also be evaluated using any hand calculator.
PROCEDURE 5.1: GIVEN ζ AND ωn
MATLAB or Calculator
1. Calculate the angle theta of the loop gain function evaluated at the desired location
scl, and subtract the angle from π using a hand calculator or the MATLAB commands
evalfr and angle.
2. Calculate the zero location using equation (5.14) using tan(theta), where theta is in radians.
3. Calculate the magnitude of the numerator of the new loop gain function, including
the controller zero, using the commands abs and evalfr, then calculate the gain using
the magnitude condition.
4. Check the time response of the PD-compensated system, and modify the design to
meet the desired specifications if necessary. Most currently available calculators
cannot perform this step.
The following MATLAB function calculates the gain and zero location for PD control.
% L is the open loop gain
% zeta and wn specify the desired closed-loop pole
% scl is the closed-loop pole, theta is the controller angle at scl
% k (a) are the corresponding gain (zero location)
function [k, a, scl] 5 pdcon(zeta, wn, L)
scl 5 wn exp(j (pi-acos(zeta))); % Find the desired closed-loop
% pole location.
theta 5 pi - angle(evalfr(L, scl)) ; % Calculate the controller
% angle.
a 5 wn sqrt(1-zeta^2)/ tan(theta) 1 zeta wn; % Calculate the
% controller zero.
Lcomp 5 L tf([1, a],1) ; % Include the controller zero.
k 5 1/abs( evalfr(Lcomp, scl)); % Calculate the gain that yields the
% desired pole.
For a specified steady-state error, the system gain is fixed and the zero location is
varied. Other design specifications require varying parameters other than the gain K.
Root locus design with a free parameter other than the gain is performed using
Procedure 5.2.
PROCEDURE 5.2: GIVEN STEADY-STATE ERROR AND ζ
1. Obtain the error constant from the steady-state error, and determine a system parameter that remains free after the error constant is fixed for the system with PD control.
2. Rewrite the closed-loop characteristic equation of the PD-controlled system in the form
1 1 Kf Gf ðsÞ 5 0
(5.15)
where Kf is a gain dependent on the free system parameter and Gf(s) is a function
of s.
3. Obtain the value of the free parameter Kf corresponding to the desired closedloop pole location using the MATLAB command rlocus. As in Procedure 5.1,
Kf can be obtained by applying the magnitude condition using MATLAB or a
calculator.
5.4 Root locus design
4. Calculate the free parameter from the gain Kf.
5. Check the time response of the PD-compensated system, and modify the design to
meet the desired specifications if necessary.
EXAMPLE 5.5
Using a CAD package, design a PD controller for the type 1 position control system of
Example 3.3 with transfer function
GðsÞ 5
1
sðs 1 4Þ
to meet the following specifications:
1. ζ 5 0.7 and ωn 5 10 rad/s
2. ζ 5 0.7 and 4% steady-state error due to a unit ramp input.
Solution
1. We solve the problem using Procedure 5.1 and the MATLAB function pdcon. Figure 5.8
shows the root locus of the uncompensated system with the desired pole location at the
intersection of the radial line for a damping ratio of 0.7 and the circular arc for an
undamped natural frequency of 10. A compensator angle of 67.2 is obtained using
(5.13) with a hand calculator or MATLAB. The MATLAB function pdcon gives
.. [k, a, scl] 5 pdcon(0.7, 10,tf(1, [1,4,0]))
k5
10.0000
a5
10.0000
scl 5
27.0000 1 7.1414i
8
0.7
7
Imaginary Axis
6
5
4
3
2
1
0
–8
–7
–6
–5
FIGURE 5.8
Root locus plot of uncompensated systems.
–4
–3
Real Axis
–2
–1
0
145
CHAPTER 5 Analog Control System Design
8
0.7
System: gcomp
Gain: 10
Pole: –7.01 + 7.13i
Damping: 0.701
Overshoot (%): 4.56
Frequency (rad/sec): 10
6
4
Imaginary Axis
146
2
10
0
–2
–4
–6
–8
–20
–18
–16
–14
0.7
–12 –10 –8
–6
Real Axis
–4
–2
0
FIGURE 5.9
Root locus plot of PD-compensated systems.
Figure 5.9 shows the root locus of the compensated plot with the cursor at the desired
pole location and a corresponding gain of 10. The results are approximately equal to
those obtained analytically in Example 5.4.
2. The specified steady-state error gives
Kv 5
100
100
Ka
5
5 25 5
.Ka 5 100
eðNÞ%
4%
4
The closed-loop characteristic equation of the PD-compensated system is given by
11K
s1a
50
sðs 1 4Þ
Let K vary with a so that their product Ka remains equal to 100, then Procedure 5.2
requires that the characteristic equation be rewritten as
11K
s
50
s2 1 4s 1 100
The corresponding root locus is a circle centered at the origin as shown in Figure 5.10
with the cursor at the location corresponding to the desired damping ratio. The desired
location is at the intersection of the root locus with the ζ 5 0.7 radial line. The corresponding gain value is K 5 10, which yields a 5 10—that is, the same values as in Example 5.3.
We obtain the value of K using the MATLAB commands
.. g 5 tfð½1; 0; ½1; 4; 100Þ; rlocusðgÞ
(Click on the root locus and drag the mouse until the desired gain is obtained.)
5.4 Root locus design
System: g
Gain: 10
Pole: –6.99 + 7.13i
Damping: 0.7
Overshoot (%): 4.59
Frequency (rad/sec): 9.98
10
8
6
Imaginary Axis
4
2
0
–2
–4
–6
–8
–10
–15
–10
–5
0
Real Axis
FIGURE 5.10
Root locus plot of PD-compensated systems with Ka fixed.
The time responses of the two designs are identical and were obtained earlier as the
cascade-compensated responses of Figures 5.6 and 5.7, respectively.
5.4.3 PI control
Increasing the type number of the system drastically improves its steady-state
response. If an integral controller is added to the system, its type number is
increased by one, but its transient response deteriorates or the system becomes
unstable. If a proportional control term is added to the integral control, the controller has a pole and a zero. The transfer function of the proportional-integral (PI)
controller is
CðsÞ 5 Kp 1
Ki
s1a
5 Kp
s
s
(5.16)
a 5 Ki =Kp
and is used in cascade compensation. An integral term in the feedback path is
equivalent to a differentiator in the forward path and is therefore undesirable (see
Problem 5.6).
147
148
CHAPTER 5 Analog Control System Design
PI design for a plant transfer function G(s) can be viewed as PD design for
the plant G(s)/s. Thus, Procedure 5.1 or 5.2 can be used for PI design. However,
a better design is often possible by placing the controller zero close to the pole
at the origin so that the controller pole and zero “almost cancel.” An almost
canceling pole-zero pair has a negligible effect on the time response. Thus, the
PI controller results in a small deterioration in the transient response with a significant improvement in the steady-state error. The following procedure can be used
for PI controller design.
PROCEDURE 5.3
1. Design a proportional controller for the system to meet the transient response specifications (i.e., place the dominant closed-loop system poles at a desired location
scl 5 2ζωn 6 jωd.
2. Add a PI controller with the zero location specified by
a5
ζ1
ωn
pffiffiffiffiffiffiffiffiffiffiffiffiffi
1 2 ζ 2 =tanðφÞ
(5.17)
or
a5
ξωn
10
(5.18)
where φ is a small angle (3-5 ).
3. Tune the gain of the system to move the closed-loop pole closer to scl.
4. Check the system time response.
If a PI controller is used, it is implicitly assumed that proportional control
meets the transient response but not the steady-state error specifications. Failure
of the first step in Procedure 5.3 indicates that a different controller (one that
improves both transient and steady-state behavior) must be used.
To prove (5.17) and justify (5.18), we use the pole-zero diagram of Figure 5.11.
The figure shows the angle contribution of the controller at the closed-loop pole
location scl. The contribution of the open-loop gain L(s) is not needed and is not
shown.
jω
scl
ωd
θp = ∠scl
θz = ∠ (scl + a)
a
ζωn
FIGURE 5.11
Pole-zero diagram of a PI controller.
×
σ
5.4 Root locus design
PROOF
The controller angle at scl is
2φ 5 θz 2 θp 5 ð180 2 θp Þ 2 ð180 2 θz Þ
From Figure 5.11, the tangents of the two angles in the preceding equation are
pffiffiffiffiffiffiffiffiffiffiffiffiffi
1 2 ζ2
ωd
tanð180 2 θp Þ5
5
ζωn
ζ
pffiffiffiffiffiffiffiffiffiffiffiffiffi
1 2 ζ2
ωd
5
tanð180 2 θz Þ 5
ζωn 2 a
ζ 2x
where x is the ratio (a/ωn). Next, we use the trigonometric identity
tanðA 2 BÞ 5
to obtain
pffiffiffiffiffiffiffiffiffi2
tanðφÞ 5
12ζ
ζ 2x
tanðAÞ 2 tanðBÞ
1 1 tanðAÞtanðBÞ
2
11
pffiffiffiffiffiffiffiffiffi2
12ζ
ζ
12ζ
ζðζ 2 xÞ
2
5
pffiffiffiffiffiffiffiffiffiffiffiffiffi
x 1 2 ζ2
1 2 ζx
Solving for x, we have
x5
ζ1
1
pffiffiffiffiffiffiffiffiffiffiffiffiffi
1 2 ζ 2 =tanðφÞ
Multiplying by ωn gives (5.17).
If the controller zero is chosen using (5.18), then x 5 ζ/10. Solving for φ we obtain
pffiffiffiffiffiffiffiffiffiffiffiffiffi!
ζ 1 2 ζ2
φ 5 tan21
10 2 ζ 2
This yields the plot of Figure 5.12, which clearly shows an angle φ of 3 or less.
φ
3.5
3
2.5
2
1.5
1
0.5
0.1
ζ
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
FIGURE 5.12
Plot of the controller angle φ at the underdamped closed-loop pole versus ζ.
149
CHAPTER 5 Analog Control System Design
The use of Procedure 5.1 or Procedure 5.3 to design PI controllers is demonstrated in Example 5.6.
EXAMPLE 5.6
Design a controller for the position control system
GðsÞ 5
1
sðs 1 10Þ
to perfectly track a ramp input and have a dominant pair with a damping ratio of 0.7 and
an undamped natural frequency of 4 rad/s.
Solution
Design 1
Apply Procedure 5.1 to the modified plant
Gi ðsÞ 5
1
s2 ðs 1 10Þ
This plant is unstable for all gains, as seen from its root locus plot in Figure 5.13. The controller must provide an angle of about 111 at the desired closed-loop pole location. Substituting
in (5.14) gives a zero at 21.732. Then moving the cursor to the desired pole location on the
root locus of the compensated system (Figure 5.14) gives a gain of about 40.6.
The design can also be obtained analytically by writing the closed-loop characteristic
polynomial as
s3 1 10s2 1 Ks 1 Ka 5 ðs 1 αÞðs2 1 2ζωn s 1 ω2n Þ
5 s3 1ðα 1 2ζωn Þs2 1ð2ζωn α 1 ω2n Þs 1 αω2n
6
5
Imaginary Axis
150
4
0.7
3
2
1
0
–10
–8
FIGURE 5.13
Root locus of a system with an integrator.
4
–6
–4
Real Axis
–2
0
5.4 Root locus design
5
4.5
System: l
Gain: 40.7
Pole: –2.8 + 2.87i
Damping: 0.698
Overshoot (%): 4.69
Frequency (rad/sec): 4.01
4
Imaginary Axis
3.5
3
2.5
2
1.5
1
0.5
0
–5 –4.5 –4 –3.5 –3 –2.5 –2 –1.5 –1 –0.5
Real Axis
0
FIGURE 5.14
Root locus of a PI-compensated system.
Then equating coefficients gives
α 5 10 2 2ζωn 5 10 2 2ð0:7Þð4Þ 5 4:4
K 5 ωn ð2ζα 1 ωn Þ 5 4 3 ½2ð0:7Þð4:4Þ 1 4 5 40:64
αω2n
4:4 3 42
a5
5
5 1:732
K
40:64
which are approximately the same as the values obtained earlier.
The MATLAB commands to obtain the zero location and the corresponding gain are
.. g 5 tfð1; ½1; 10; 0; 0Þ; scl 5 4 ð20:7 1 j sqrtð1 2 0:7^2ÞÞ
scl 5 22:8000 1 2:8566i
.. theta 5 pi 2 angleðpolyvalð½1; 10; 0; 0;sclÞÞ
theta 5 1:9285
.. a 5 imagðsclÞ=tanðthetaÞ 2 realðsclÞ
a 5 1:7323
.. k 5 1=absðevalfrðg tfð½1; a; 1Þ; sclÞÞ
k 5 40:6400
The closed-loop transfer function for the preceding design (Design 1) is
Gcl ðsÞ 5
5
40:64ðs 1 1:732Þ
s3 1 10s2 1 40:64s 1 69:28
40:64ðs 1 1:732Þ
ðs 1 4:4Þðs2 1 5:6s 1 16Þ
The system has a zero close to the closed-loop poles, which results in excessive overshoot
in the time response of Figure 5.15 (shown together with the response for an alternative
design).
151
CHAPTER 5 Analog Control System Design
1
Output
152
0.5
0
0
1
2
3
4
5
Time s
FIGURE 5.15
Step response of a PI-compensated system: Design 1 (dotted) and Design 2 (solid).
Design 2
Next, we apply Procedure 5.3 to the same problem. The proportional control gain for a
damping ratio of 0.7 is approximately 51.02 and yields an undamped natural frequency of
7.143 rad/s. This is a faster design than required and is therefore acceptable. Then we use
(5.17) and obtain the zero location
a5
5
ζ1
ωn
pffiffiffiffiffiffiffiffiffiffiffiffiffi
1 2 ζ 2 =tanð2φÞ
0:7 1
7:143
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
D0:5
1 2 0:49=tanð3 Þ
If (5.18) is used, we have
a5
ζωn
7:143 3 0:7
5
D 0:5
10
10
That is, the same zero value is obtained.
The closed-loop transfer function for this design is
Gcl ðsÞ 5
5
50ðs 1 0:5Þ
s3 1 10s2 1 50s 1 25
50ðs 1 0:5Þ
ðs 1 0:559Þðs2 1 9:441s 1 44:7225Þ
where the gain value has been slightly reduced to bring the damping ratio closer to 0.7.
This actually does give a damping ratio of about 0.7 and an undamped natural frequency
of 6.6875 rad/s for the dominant closed-loop poles.
5.4 Root locus design
The time response for this design (Design 2) is shown, together with that of Design 1,
in Figure 5.15. The percentage overshoot for Design 2 is much smaller because its zero
almost cancels with a closed-loop pole. Design 2 is clearly superior to Design 1.
5.4.4 PID control
If both the transient and steady-state response of the system must be improved,
then neither a PI nor a PD controller may meet the desired specifications. Adding
a zero (PD) may improve the transient response but does not increase the type
number of the system. Adding a pole at the origin increases the type number but
may yield an unsatisfactory time response even if one zero is also added. With a
proportional-integral-derivative (PID) controller, two zeros and a pole at the
origin are added. This both increases the type number and allows satisfactory
reshaping of the root locus.
The transfer function of a PID controller is given by
Ki
s2 1 2ζωn s 1 ω2n
1 Kd s 5 Kd
s
s
2
2ζωn 5 Kp =Kd ; ωn 5 Ki =Kd
CðsÞ 5 Kp 1
(5.19)
where Kp, Ki, and Kd are the proportional, integral, and derivative gain, respectively.
The zeros of the controller can be real or complex conjugate, allowing the cancellation of real or complex conjugate LHP poles if necessary. In some cases, good
design can be obtained by canceling the pole closest to the imaginary axis. The
design then reduces to a PI controller design that can be completed by applying
Procedure 5.3. Alternatively, one could apply Procedure 5.1 or 5.2 to the reduced
transfer function with an added pole at the origin. A third approach to PID design
is to follow Procedure 5.3 with the proportional control design step modified to
PD design. The PD design is completed using Procedure 5.1 or 5.2 to meet the
transient response specifications. PI control is then added to improve the steady-state
response. Examples 5.7 and 5.8 illustrate these design procedures.
EXAMPLE 5.7
Design a PID controller for an armature-controlled DC motor with transfer function
GðsÞ 5
1
sðs 1 1Þðs 1 10Þ
to obtain zero steady-state error due to ramp, a damping ratio of 0.7, and an undamped
natural frequency of 4 rad/s.
Solution
Canceling the pole at 21 with a zero and adding an integrator yields the transfer function
Gi ðsÞ 5
1
s2 ðs 1 10Þ
153
CHAPTER 5 Analog Control System Design
This is identical to the transfer function obtained in Example 5.6. Hence, the overall
PID controller is given by
CðsÞ 5 50
ðs 1 1Þðs 1 0:5Þ
s
This design is henceforth referred to as Design 1.
A second design (Design 2) is obtained by first selecting a PD controller to meet the
transient response specifications. We seek an undamped natural frequency of 5 rad/s
in anticipation of the effect of adding PI control. The PD controller is designed using the
MATLAB commands (using the function pdcon)
.. ½k; a; scl 5 pdconð0:7; 5;tfð1; ½1; 11; 10; 0ÞÞ
k 5 43:0000
a 5 2:3256
scl 5 23:5000 1 3:5707i
The PI zero is obtained using (5.17) and the command
.. b 5 5=ð0:7 1 sqrtð1 2 :49Þ=tanð3 pi=180ÞÞ
b 5 0:3490
Note that this results is almost identical to the zero obtained location obtained more easily
using (5.18). For a better transient response, the gain is reduced to 40 and the controller
transfer function for Design 2 is
CðsÞ 5 40
ðs 1 0:349Þðs 1 2:326Þ
s
The step responses for Designs 1 and 2 are shown in Figure 5.16. Clearly, Design 1 is
superior because the zeros in Design 2 result in excessive overshoot. The plant transfer
1.4
1.2
1
Output
154
0.8
0.6
0.4
0.2
0
0
1
2
3
Time s
FIGURE 5.16
Time response for Design 1 (dotted) and Design 2 (solid).
4
5
5.4 Root locus design
function favors pole cancellation in this case because the remaining real axis pole is far
in the LHP. If the remaining pole is at, say, 23, the second design procedure would
give better results. The lesson to be learned is that there are no easy solutions in design.
There are recipes with which the designer should experiment until satisfactory results are
obtained.
EXAMPLE 5.8
Design a PID controller to obtain zero steady-state error due to step, a damping ratio of 0.7,
and an undamped natural frequency of at least 4 rad/s for the transfer function
GðsÞ 5
1
ðs 1 10Þðs2 1 2s 1 10Þ
Solution
The system has a pair of complex conjugate poles that slow down its time response and a
third pole that is far in the LHP. Canceling the complex conjugate poles with zeros and
adding the integrator yields the transfer function
GðsÞ 5
1
sðs 1 10Þ
The root locus of the system is similar to Figure 5.2(a), and we can increase the gain
without fear of instability. The closed-loop characteristic equation of the compensated
system with gain K is
s2 1 10s 1 K 5 0
1.2
1
Output
0.8
0.6
0.4
0.2
0
0
0.5
1
Time s
FIGURE 5.17
Step response of the PID-compensated system of Example 5.8.
1.5
155
156
CHAPTER 5 Analog Control System Design
Equating coefficients as in Example 5.3, we observe that for a damping ratio of 0.7
the undamped natural frequency is
ωn 5
10
5
5
5 7:143 rad=s
2ζ
0:7
This meets the design specifications. The corresponding gain is 51.02, and the PID
controller is given by
CðsÞ 5 51:02
s2 1 2s 1 10
s
In practice, pole-zero cancellation may not occur, but near cancellation is sufficient to
obtain a satisfactory time response, as shown in Figure 5.17.
5.5 Empirical tuning of PID controllers
In industrial applications, PID controllers are often tuned empirically. Typically,
the controller parameters are selected based on a simple process model using
a suitable tuning rule. This allows us to address (1) load disturbance rejection
specifications (which are often a major concern in process control) and (2) the
presence of a time delay in the process. We first write the PID controller transfer
function (5.19) in the form
1
1 Td s
CðsÞ 5 Kp 1 1
Ti s
Ti 5 Kp =Ki ;
(5.20)
Td 5 Kd =Kp
where Ti denotes the integral time constant and Td denotes the derivative time
constant. The three controller parameters Kp, Ti, and Td have a clear physical
meaning. Increasing Kp (i.e., increasing the proportional action) provides a faster
but more oscillatory response. The same behavior results from increasing the integral action by decreasing the value of Ti. Finally, increasing the value of Td leads
to a slower but more stable response.
These considerations allow tuning the controller by a trial-and-error procedure.
However, this can be time consuming, and the achieved performance depends on
the skill of the designer. Fortunately, tuning procedures are available to simplify
the PID design. Typically, the parameters of the process model are determined
assuming a first-order-plus-dead-time model. That is,
GðsÞ 5
K
e2Ls
τs 1 1
(5.21)
5.5 Empirical tuning of PID controllers
where K is the process gain, τ is the process (dominant) time constant, and L is
the (apparent) dead time of the process. The response for the approximate model
of (5.21) due to a step input of amplitude A is given by the expression
t2L
1ðt 2 LÞ
KA 1 2 exp 2
τ
(5.22)
The rising exponential of (5.22) has a steady-state level of K A and a tangent to
it at the initiation of its rise and with the initial slope reaches the steady-state
level after one time constant τ. In addition, the rising exponential reaches 63% of
the final value after one time constant. Hence, we can estimate the parameters of
the approximate model based on the step response of the process through the
tangent method.2 The method consists of the following steps.
TANGENT METHOD
1. Obtain the step response of the process experimentally.
2. Draw a tangent to the step response at the inflection point as shown in Figure 5.18.
3. Compute the process gain as the ratio of the steady-state change in the process output y
to the amplitude of the input step A.
τ
1
Process Output
0.8
0.6
KA
0.4
0.2
0
–0.2
τ
L
0
2
4
6
8
10
12
14
Time s
FIGURE 5.18
Application of the tangent method.
2
See Åström and Hägglund (2006) and Visioli (2006) for a detailed description and analysis of
different methods for estimating the parameters of a first-order-plus-dead-time system.
157
158
CHAPTER 5 Analog Control System Design
4. Compute the apparent dead time L as the time interval between the application of the
step input and the intersection of the tangent line with the time axis.
5. Determine the sum τ 1 L (from which the value of τ can be easily computed) as the time
interval between the application of the step input and the intersection of the tangent
line with the straight representing the final steady-state value of the process output.
Alternatively, the value of τ 1 L can be determined as the time interval between the
application of the step input and the time when the process output attains 63.2% of its
final value. Note that if the dynamics of the process can be perfectly described by
a first-order-plus-dead-time model, the values of τ obtained in the two methods are
identical.
Given the process model parameters, several tuning rules are available for
determining the PID controller parameter values, but different rules address
different design specifications. The most popular tuning rules are those attributed
to Ziegler-Nichols. Their aim is to provide satisfactory load disturbance rejection.
Table 5.1 shows the Ziegler-Nichols rules for P, PI, and PID controllers.
Although the rules are empirical, they are consistent with the physical meaning
of the parameters. For example, consider the effect of the derivative action in PID
control governed by the third row of Table 5.1. The derivative action provides
added damping to the system, which increases its relative stability. This allows us
to increase both the proportional and integral action while maintaining an
acceptable time response. We demonstrate the Ziegler-Nichols procedure using
Example 5.9.
Table 5.1 Ziegler-Nichols Tuning Rules for a First-OrderPlus-Dead-Time Model of the Process
Controller Type
P
Kp
τ
KL
Ti
Td
PI
0:9
τ
KL
3L
PID
1:2
τ
KL
2L
0.5 L
EXAMPLE 5.9
Consider the control system shown in Figure 5.19, where the process has the following
transfer function:
GðsÞ 5
1
e20:2s
ðs11Þ4
5.5 Empirical tuning of PID controllers
Load
Disturbance
Reference
Input +
+
PID
Controller
−
Output
G(s)
FIGURE 5.19
Block diagram of the process of Example 5.9.
1
0.9
0.8
Process Output
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0
2
4
6
8
10
12
14
Time s
FIGURE 5.20
Application of the tangent method in Example 5.9.
Estimate a first-order-plus-dead-time model of the process, and design a PID controller by
applying the Ziegler-Nichols tuning rules.
Solution
The process step response is shown in Figure 5.20. By applying the tangent
method, a first-order-plus-dead-time model with gain K 5 1, L 5 1.55 and a delay τ 5 3
is estimated. The Ziegler-Nichols rules of Table 5.1 provide the following PID parameters: Kp 5 2.32, Ti 5 3.1, and Td 5 0.775. Figure 5.21 shows the response of the
closed-loop system due to a step input at t 5 0 followed by a step load disturbance
input at t 5 50.
159
CHAPTER 5 Analog Control System Design
1.5
Process Output
160
1
0.5
0
0
10
20
30
40
50 60
Time s
70
80
90
100
FIGURE 5.21
Process output with the PID controller tuned with the Ziegler-Nichols method.
The system response is oscillatory, which is a typical feature of the Ziegler-Nichols
method, whose main objective is to satisfy load disturbance response specifications.
However, the effect of the load step disturbance is minimized.
Ziegler and Nichols also devised tuning rules based on using the critical gain
Kc (that is, the gain margin) and the period of the corresponding oscillations Tc.
The data needed for this procedure can be obtained experimentally using proportional feedback control by progressively increasing the controller gain until the
system becomes marginally stable, and then recording the controller gain and the
period of the resulting oscillations. The controller parameters are then computed
using the rules of Table 5.2. The main drawback of this approach is that it
requires the system to operate on the verge of instability, which is often detrimental. However, this problem can be overcome using a relay in the feedback loop
to cause oscillatory behavior, known as a limit cycle.3 Controller tuning using
a closed-loop system with proportional control is explored through a computer
exercise (see Problem 5.16).
Table 5.2 Ziegler-Nichols Tuning Rules for the Closed-Loop Method
3
Controller Type
Kp
Ti
Td
P
PI
PID
0.5 Kc
0.4 Kc
0.6 Kc
_
0.8 Tc
0.5 Tc
0.125 Tc
See Åström and Hägglund (2006) for details on the relay-feedback methodology.
Problems
Resources
Åström, K.J., Hägglund, T., 2006. Advanced PID Controllers. ISA Press, Research
Triangle Park, NJ.
D’Azzo, J.J., Houpis, C.H., 1988. Linear Control System Analysis and Design. McGraw-Hill,
New York.
Kuo, B.C., 1995. Automatic Control Systems. Prentice Hall, Englewood Cliffs, NJ.
Nise, N.S., 1995. Control Systems Engineering. Benjamin Cummings, Hoboken, NJ.
O’Dwyer, A., 2006. Handbook of PI and PID Tuning Rules. Imperial College Press,
London, UK.
Ogata, K., 1990. Modern Control Engineering. Prentice Hall, Englewood Cliffs, NJ.
Van de Vegte, J., 1986. Feedback Control Systems. Prentice Hall, Englewood Cliffs, NJ.
Visioli, A., 2006. Practical PID Control. Springer, UK.
PROBLEMS
5.1
Prove that for a system with two real poles and a real zero
LðsÞ 5
s1a
;
ðs 1 p1 Þðs 1 p2 Þ
a , p1 , p2 or p1 , p2 , a
the breakaway point is at a distance
5.2
Use the result of Problem 5.1 to draw the root locus of the system
KLðsÞ 5
5.3
Kðs 1 4Þ
sðs 1 2Þ
Sketch the root loci of the following systems:
K
(a) KLðsÞ 5
sðs 1 2Þðs 1 5Þ
(b) KLðsÞ 5
5.4
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ða 2 p1 Þða 2 p2 Þ from the zero.
Kðs 1 2Þ
sðs 1 3Þðs 1 5Þ
Consider the system in 5.3(b) with a required steady-state error of 20%,
and an adjustable PI controller zero location. Show that the corresponding
closed-loop characteristic equation is given by
11K
s1a
1
50
s
ðs 1 3Þðs 1 5Þ
Next, rewrite the equation as
1 1 Kf Gf ðsÞ 5 0
161
162
CHAPTER 5 Analog Control System Design
where Kf 5 K, K.a is constant, and Gf (s) is a function of s, and examine
the effect of shifting the zero on the closed-loop poles.
(a) Design the system for a dominant second-order pair with a damping
ratio of 0.5. What is ωn for this design?
(b) Obtain the time response using a CAD program. How does the time
response compare with that of a second-order system with the same ωn
and ζ as the dominant pair? Give reasons for the differences.
(c) Discuss briefly the trade-off between error, speed of response, and
relative stability in this problem.
5.5
Prove equations (5.13) and (5.14), and justify the design Procedures 5.1 and 5.2.
5.6
Show that a PI feedback controller is undesirable because it results in
a differentiator in the forward path. Discuss the step response of the closedloop system.
5.7
Design a controller for the transfer function
GðsÞ 5
1
ðs 1 1Þðs 1 5Þ
to obtain (a) zero steady-state error due to step, (b) a settling time of less
than 2 s, and (c) an undamped natural frequency of 5 rad/s. Obtain the
response due to a unit step, and find the percentage overshoot, the time to
the first peak, and the steady-state error percentage due to a ramp input.
5.8
Repeat Problem 5.7 with a required settling time less than 0.5 s and an
undamped natural frequency of 10 rad/s.
5.9
Consider the oven temperature control system of Example 3.5 with transfer
function
GðsÞ 5
K
s2 1 3s 1 10
(a) Design a proportional controller for the system to obtain a percentage
overshoot less than 5%.
(b) Design a controller for the system to reduce the steady-state error
due to step to zero without significant deterioration in the transient
response.
5.10
For the inertial system governed by the differential equation
θ€ 5 τ
design a feedback controller to stabilize the system and reduce the percentage overshoot below 10% with a settling time of less than 4 s.
Computer Exercises
COMPUTER EXERCISES
5.11
Consider the oven temperature control system described in Example 3.5
with transfer function
GðsÞ 5
s2
K
1 3s 1 10
(a) Obtain the step response of the system with a PD cascade controller
with gain 80 and a zero at 25.
(b) Obtain the step response of the system with PD feedback controller
with a zero at 25 and unity gain and forward gain of 80.
(c) Why is the root locus identical for both systems?
(d) Why are the time responses different although the systems have the
same loop gains?
(e) Complete a comparison table using the responses of (a) and (b),
including the percentage overshoot, the time to first peak, the settling
time, and the steady-state error. Comment on the results, and explain
the reason for the differences in the response.
5.12
Use Simulink to examine a practical implementation of the cascade
controller described in Exercise 5.11. The compensator transfer function
includes a pole because PD control is only approximately realizable. The
controller transfer is of the form
CðsÞ 5 80
0:2s 1 1
0:02s 1 1
(a) Simulate the system with a step reference input both with and without
a saturation block with saturation limits 65 between the controller
and plant. Export the output to MATLAB for plotting (you can use a
Scope block and select “Save data to workspace”).
(b) Plot the output of the system with and without saturation together, and
comment on the difference between the two step responses.
5.13
Consider the system
GðsÞ 5
1
ðs11Þ4
and apply the Ziegler-Nichols procedure based on Table 5.1 to design
a PID controller. Obtain the response due to a unit step input as well as a
unit step disturbance signal.
5.14
Write a computer program that implements the estimation of a first-orderplus-dead-time transfer function with the tangent method and then determine
163
164
CHAPTER 5 Analog Control System Design
the PID parameters using the Ziegler-Nichols formula. Apply the program
to the system
GðsÞ 5
1
ðs11Þ8
and simulate the response of the control system when a set-point step
change and a load disturbance step are applied. Discuss the choice of the
time constant value based on the results.
5.15
Apply the script of Exercise 5.14 to the system
GðsÞ 5
1
ðs11Þ2
and simulate the response of the control system when a set-point step
change and a load disturbance step are applied. Compare the results
obtained with those of Problem 5.14.
5.16
Use the Ziegler-Nichols closed-loop method to design a PID controller for
the system
GðsÞ 5
1
ðs11Þ4
based on Table 5.2. Obtain the response due to a unit step input together
with a unit step disturbance signal. Compare the results with those of
Problem 5.13.
CHAPTER
Digital Control System Design
6
OBJECTIVES
After completing this chapter, the reader will be able to do the following:
1. Sketch the z-domain root locus for a digital control system, or obtain it using
MATLAB.
2. Obtain and tune a digital controller from an analog design.
3. Design a digital controller in the z-domain directly using the root locus approach.
4. Design a digital controller directly using frequency domain techniques.
5. Design a digital controller directly using the synthesis approach of Ragazzini.
6. Design a digital control system with finite settling time.
To design a digital control system, we seek a z-domain transfer function or difference equation model of the controller that meets given design specifications. The
controller model can be obtained from the model of an analog controller that
meets the same design specifications. Alternatively, the digital controller can be
designed in the z-domain using procedures that are almost identical to s-domain
analog controller design. We discuss both approaches in this chapter. We begin
by introducing the z-domain root locus.
6.1 z-Domain root locus
In Chapter 3, we showed that the closed-loop characteristic equation of a digital
control system is of the form
1 1 CðzÞGZAS ðzÞ 5 0
(6.1)
where C(z) is the controller transfer function and GZAS(z) is the transfer function
of the DAC, analog subsystem, and ADC combination. If the controller is
assumed to include a constant gain multiplied by a rational z-transfer function,
then (6.1) is equivalent to
1 1 KLðzÞ 5 0
(6.2)
where L(z) is the open-loop gain.
Equation (6.2) is identical in form to the s-domain characteristic equation
(5.1) with the variable s replaced by z. Thus, all the rules derived for equation
165
CHAPTER 6 Digital Control System Design
(5.1) are applicable to (6.2) and can be used to obtain z-domain root locus plots.
The plots can also be obtained using the root locus plots of most computer-aided
design (CAD) programs. Thus, we can use the MATLAB command rlocus for
z-domain root loci.
EXAMPLE 6.1
Obtain the root locus plot and the critical gain for the first-order type 1 system with loop
gain
LðzÞ 5
1
z21
Solution
Using root locus rules gives the root locus plot in Figure 6.1, which can be obtained
using the MATLAB command rlocus. The root locus lies entirely on the real axis between
the open-loop pole and the open-loop zero. For a stable discrete system, real axis z-plane
poles must lie between the point (21, 0) and the point (1, 0). The critical gain for the
system corresponds to the point (21, 0). The closed-loop characteristic equation of the
system is
z211K 50
Substituting z 5 21 gives the critical gain Kcr 5 2, as shown on the root locus plot.
1
0.8
0.6
0.4
Imaginary Axis
166
0.2
System: g
Gain: 2
Pole: –1
Damping: 0.000141
Overshoot (%): 100
Frequency (rad/sec): 314
0
–0.2
–0.4
–0.6
–0.8
–1
–1.5
–1
FIGURE 6.1
Root locus of a type 1 first-order system.
–0.5
0
Real Axis
0.5
1
6.1 z-Domain root locus
EXAMPLE 6.2
Obtain the root locus plot and the critical gain for the second-order type 1 system with
loop gain
LðzÞ 5
1
ðz 2 1Þðz 2 0:5Þ
Solution
Using root locus rules gives the root locus plot in Figure 6.2, which has the same form as
the root locus of Example 5.1 (1) but is entirely in the right-hand plane (RHP). The breakaway point is midway between the two open-loop poles at zb 5 0.75. The critical gain now
occurs at the intersection of the root locus with the unit circle. To obtain the critical gain
value, first write the closed-loop characteristic equation
ðz 2 1Þðz 2 0:5Þ 1 K 5 z2 2 1:5 z 1 K 1 0:5 5 0
On the unit circle, the closed-loop poles are complex conjugate and of magnitude unity.
Hence, the magnitude of the poles satisfies the equation
2
z1;2 5 Kcr 1 0:5 5 1
where Kcr is the critical gain. The critical gain is equal to 0.5, which, from the closed-loop
characteristic equation, corresponds to unit circle poles at
z1;2 5 0:75 6 j0:661
System: g
Gain: 0.5
Pole: 0.75 + 0.661i
Damping: –4.84e – 005
Overshoot (%): 100
Frequency (rad/sec): 72.3
1
0.8
Imaginary Axis
0.6
0.4
0.2
0
–0.2
–0.4
–0.6
–0.8
–1
–1.5
–1
–0.5
0
Real Axis
FIGURE 6.2
z-root locus of a type 1 second-order system.
0.5
1
1.5
167
168
CHAPTER 6 Digital Control System Design
6.2 z-Domain digital control system design
In Chapter 5, we were able to design analog control systems by selecting their poles
and zeros in the s-domain to correspond to the desired time response. This approach
was based on the relation between any time function and its s-domain poles and
zeros. If the time function is sampled and the resulting sequence is z-transformed, the
z-transform contains information about the transformed time sequence and the original time function. The poles of the z-domain function can therefore be used to characterize the sequence, and possibly the sampled continuous time function, without
inverse z-transformation. However, this latter characterization is generally more difficult than characterization based on the s-domain functions described by Figure 5.3.
Figure 6.3 shows z-domain pole locations and the associated temporal sequences.
As in the continuous case, positive real poles are associated with exponentials.
Unlike the continuous case, the exponentials decay for poles inside the unit circle
and increase for poles outside it. In addition, negative poles are associated with
sequences of alternating signs. Poles on the unit circle are associated with a response
of constant magnitude. For complex conjugate poles, the response is oscillatory, with
the rate of decay determined by the pole distance from the origin and the frequency
of oscillations determined by the magnitude of the pole angle. Complex conjugate
poles on the unit circle are associated with sustained oscillations.
Because it is easier to characterize the time functions using s-domain poles,
it may be helpful to reexamine Figure 5.3 and compare it to Figure 6.3. Comparing
the figures suggests a relationship between s-domain and z-domain poles that
greatly simplifies z-domain pole characterization. To obtain the desired relationship,
we examine two key cases, both in the s-domain and in the z-domain. One case
yields a real pole and the other a complex conjugate pair of poles. More complex
time functions can be reduced to these cases by partial fraction expansion of the
transforms. The two cases are summarized in Tables 6.1 and 6.2.
Im{z}
×
×
×
×
×
×
×
×
FIGURE 6.3
z-domain pole locations and the associated temporal sequences.
×
Re{z}
Table 6.1 Time Functions and Real Poles
Laplace Transform
t$0
t,0
FðsÞ 5
1
s1α
Sampled
2αkT
; k$0
e
f ðkTÞ 5
0;
k,0
z-Transform
FðzÞ 5
z
z 2 e2αT
Table 6.2 Time Functions and Complex Conjugate Poles
Continuous
2αt
e sinðωd tÞ;
f ðtÞ 5
0;
Laplace Transform
t$0
t,0
FðsÞ 5
ωd
ðs1αÞ2 1 ω2d
Sampled
2αkT
e
sinðωd kTÞ;
f ðkTÞ 5
0;
z-Transform
k$0
k,0
FðzÞ 5
sinðωd TÞe2αT z
z2 2 2cosðωd TÞe2αT 1 e22αT
6.2 z-Domain digital control system design
Continuous
2αt
e ;
f ðtÞ 5
0;
169
170
CHAPTER 6 Digital Control System Design
Using the two tables, it appears that if F(s) has a pole at 2α, F(z) has a pole
at e2αT, and if F(s) has poles at 2ζωn 1 jωd, F(z) has poles at eð2ζωn 1jωd ÞT . We
therefore make the following observation.
Observation
If the Laplace transform F(s) of a continuous-time function f(t) has a pole ps, then
the z-transform F(z) of its sampled counterpart f(kT) has a pole at
pz 5 eps T
(6.3)
where T is the sampling period.
REMARKS
1. The preceding observation is valid for a unit step with its s-domain pole at the origin
because the z-transform of a sampled step has a pole at 1 5 e0.
2. There is no general mapping of s-domain zeros to z-domain zeros.
3. From (6.3), the z-domain poles in the complex conjugate case are given by
pz 5 eσT ejωd T
5 eσT ejðωd T1k2πÞ ;
k 5 0; 1; 2; . . .
Thus, pole locations are a periodic function of the damped natural frequency
ωd with period (2π/T) (i.e., the sampling angular frequency ωs). The mapping of
distinct s-domain poles to the same z-domain location is clearly undesirable in
situations where a sampled waveform is used to represent its continuous counterpart. The strip of width ωs over which no such ambiguity occurs (frequencies in
the range [(2ωs /2), ωs /2] rad/s) is known as the primary strip (Figure 6.4). The
width of this strip can clearly be increased by faster sampling, with the choice of
suitable sampling rate dictated by the nature of the continuous time function. We
observe that the minimum sampling frequency for good correlation between the
s
2
σ
Primary
Strip
−
FIGURE 6.4
Primary strip in the s-plane.
s
2
6.2 z-Domain digital control system design
analog and digital signals is twice the frequency ωd as expected based on the sampling theorem of Section 2.9.
6.2.1 z-Domain contours
Using the observation (6.3), one can use s-domain contours over which certain
characteristics of the function poles are fixed to obtain the shapes of similar
z-domain contours. In particular, the important case of a second-order underdamped system yields Table 6.3.
The information in the table is shown in Figures 6.5 and 6.6 and can be used to
predict the time responses of Figure 6.3. From Figure 6.5, we see that negative
values of σ correspond to the inside of the unit circle in the z-plane, whereas positive values correspond to the outside of the unit circle. The unit circle is the σ 5 0
contour. Both Table 6.3 and Figure 6.5 also show that large positive σ values
correspond to circuits of large radii, whereas large negative σ values correspond
to circles of small radii. In particular, a positive infinite σ corresponds to the point
at N and a negative infinite σ corresponds to the origin of the z-plane.
Table 6.3 Pole Contours in the s-Domain and the z-Domain
Contour
s-Domain Poles
σ 5 constant
ωd 5 constant
Contour
z-Domain Poles
σT
jzj 5 e 5 constant
+z 5 ωdT 5 constant
Vertical line
Horizontal line
Circle
Radial line
Im(z)
3
j
5
4
2
3
2
10
6
2
1
–2
–6
–10
1
0
0
–1
–1
–2
–3
–2
–4
–5
–10
–5
0
(a)
5
10 σ
–3
–3
–2
–1
0
(b)
1
2
3 Re(z)
FIGURE 6.5
Constant σ contours in the s-plane and in the z-plane. (a) Constant σ contours in the
s-plane. (b) Constant σ contours in the z-plane.
171
172
CHAPTER 6 Digital Control System Design
Im(z)
3
j
3 /4
3 /4
2
/2
0
0
− /4
–2
−3 /4
–1
0
(a)
1
2
−
0
–1
− /2
–2
/4
1
/4
–3
/2
3 σ
− /2
−3 /4
–3
–3
–2
–1
0
(b)
1
− /4
2
3 Re(z)
FIGURE 6.6
Constant ωd contours in the s-plane and the z-plane. (a) Constant ωd contours in the
s-plane. (b) Constant ωd contours in the z-plane.
From Figure 6.6, we see that larger ωd values correspond to larger angles,
with ωd 5 6ωs/2 corresponding to 6π. As observed earlier using a different
argument, a system with poles outside this range (outside the primary strip)
does not have a one-to-one correspondence between s-domain and z-domain
poles.
The z-domain characteristic polynomial for a second-order underdamped system is
ðz 2 eð2ζωn 1jωd ÞT Þðz 2 eð2ζωn 2jωd ÞT Þ 5 z2 2 2cosðωd TÞe2ζωn T z 1 e22ζωn T
Hence, the poles of the system are given by
z1;2 5 e2ζωn T +6ωd T
(6.4)
This confirms that constant ζωn contours are circles, whereas constant ωd contours are radial lines.
Constant ζ lines are logarithmic spirals that get smaller for larger values of ζ.
The spirals are defined by the equation
jzj 5 e
p2ζθ
ffiffiffiffiffiffi
12ζ 2
2ζ ðπθ3 =1803 Þ
5e
pffiffiffiffiffiffi
12ζ 2
(6.5)
where jzj is the magnitude of the pole and θ is its angle. Constant ωn contours are
defined by the equation
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ffi
2
2
jzj 5 e2 ðωn TÞ 2θ
(6.6)
Figure 6.7 shows constant ζ contours and constant ωn contours.
6.2 z-Domain digital control system design
0.5 π/T
0.4 π/T
0.6 π/T
1
0.7 π/T
0.8
0.8 π/T
0.6
0.4
0.3
0.9 π/T
π/T
0.2 π/T
0.1 π/T
π/T
–0.2
–0.4
0.2
0.4
0.5
0.6
0.7
0.8
0.9
0.2
0
0.3 π/T
0.1
0.1 π/T
0.9 π/T
–0.6
0.8 π/T
–0.8
0.2 π/T
0.7 π/T
0.3 π/T
0.6 π/T
–1
–1
0.5 π/T
–0.5
0
0.4 π/T
0.5
1
FIGURE 6.7
Constant ζ and ωn contours in the z-plane.
To prove (6.5), rewrite the complex conjugate poles as
z1;2 5 e2ζωn T +6ωd T 5 jzj+6θ
or equivalently,
θ 5 ωd T 5 ωn T
qffiffiffiffiffiffiffiffiffiffiffiffiffi
1 2 ζ2
jzj 5 e2ζωn T
(6.7)
(6.8)
(6.9)
Eliminating ωn T from (6.9) using (6.8), we obtain the spiral equation (6.5).
The proof of (6.6) is similar and is left as an exercise (Problem 6.2).
The following observations can be made by examining the spiral equation
(6.5):
1. For every ζ value, there are two spirals corresponding to negative and positive
angles θ. The negative θ spiral is below the real axis and is the mirror image
of the positive θ spiral.
2. For a given spiral, the magnitude of the pole drops logarithmically with its
angle.
3. At the same angle θ, increasing the damping ratio gives smaller pole
magnitudes. Hence, the spirals are smaller for larger ζ values.
4. All spirals start at θ 5 0, jzj 5 1 but end at different points.
173
174
CHAPTER 6 Digital Control System Design
5. For a given damping ratio and angle θ, the pole magnitude can be obtained by
substituting in (6.5). For a given damping ratio and pole magnitude, the pole
angle can be obtained by substituting in the equation
pffiffiffiffiffiffiffiffiffiffiffiffiffi
1 2 ζ 2 lnðjzjÞ
θ5
ζ
(6.10)
The standard contours can all be plotted using a handheld calculator. Their
intersection with the root locus can then be used to obtain the pole locations for
desired closed-loop characteristics. However, it is more convenient to obtain the
root loci and contours using a CAD tool, especially for higher-order systems. The
unit circle and constant ζ and ωn contours can be added to root locus plots
obtained with CAD packages to provide useful information on the significance of
z-domain pole locations. In MATLAB, this is accomplished using the command
.. zgridð zeta; wnÞ
where zeta is a vector of damping ratios and wn is a vector of the undamped natural frequencies for the contours.
Clearly, the significance of pole and zero locations in the z-domain is completely
different from identical locations in the s-domain. For example, the stability boundary in the z-domain is the unit circle, not the imaginary axis. The characterization of
the time response of a discrete-time system based on z-domain information is more
complex than the analogous process for continuous-time systems based on the sdomain information discussed in Chapter 5. These are factors that slightly complicate
z-domain design, although the associated difficulties are not insurmountable.
The specifications for z-domain design are similar to those for s-domain
design. Typical design specifications are as follows:
Time constant. This is the time constant of exponential decay for the
continuous envelope of the sampled waveform. The sampled signal is
therefore not necessarily equal to a specified portion of the final value
after one time constant. The time constant is defined as
τ5
1
ζωn
(6.11)
Settling time. The settling time is defined as the period after which the
envelope of the sampled waveform stays within a specified percentage (usually
2%) of the final value. It is a multiple of the time constant depending on the
specified percentage. For a 2% specification, the settling time is given by
Ts 5
4
ζωn
(6.12)
Frequency of oscillations ωd. This frequency is equal to the angle of the
dominant complex conjugate poles divided by the sampling period.
6.2 z-Domain digital control system design
Other design criteria such as the percentage overshoot, the damping ratio, and the
undamped natural frequency can also be defined analogously to the continuous case.
6.2.2 Proportional control design in the z-domain
Proportional control involves the selection of a DC gain value that corresponds to
a time response satisfying design specifications. As in s-domain design, a satisfactory time response is obtained by tuning the gain to select a dominant closed-loop
pair in the appropriate region of the complex plane. Analytical design is possible
for low-order systems but is more difficult than its analog counterpart. Example
6.3 illustrates the design of proportional digital controllers.
EXAMPLE 6.3
Design a proportional controller for the digital system described in Example 6.2 with a
sampling period T 5 0.1 s to obtain
1. A damped natural frequency of 5 rad/s
2. A time constant of 0.5 s
3. A damping ratio of 0.7
Solution
After some preliminary calculations, the design results can be easily obtained using the
rlocus command of MATLAB. The following calculations, together with the information provided by a cursor command, allow us to determine the desired closed-loop pole locations:
1. The angle of the pole is ωdT 5 5 3 0.1 5 0.5 rad or 28.65 .
2. The reciprocal of the time constant is ζωn 5 1/0.5 5 2 rad/s. This yields a pole magnitude
of e2ζωn T 5 0:82.
3. The damping ratio given can be used directly to locate the desired pole.
Using MATLAB, we obtain the results shown in Table 6.4. The corresponding sampled
step response plots obtained using the command step (MATLAB) are shown in Figure 6.8.
As expected, the higher gain designs are associated with a low damping ratio and a more
oscillatory response.
Table 6.4 results can also be obtained analytically using the characteristic equation for
the complex conjugate poles of (6.4). The system’s closed-loop characteristic equation is
z2 2 1:5z 1 K 1 0:5 5 z2 2 2cosðωd TÞe2ζωn T z 1 e22ζωn T
Equating coefficients gives the two equations
z1 : 1:5 5 2cosðωd TÞe2ζωn T
z0 : K 1 0:5 5 e22ζωn T
Table 6.4 Proportional Control Design Results
Design
Gain
ζ
ωn rad/s
(a)
(b)
(c)
0.23
0.17
0.10
0.3
0.4
0.7
5.24
4.60
3.63
175
CHAPTER 6 Digital Control System Design
1.4
1.2
1
Output
176
0.8
0.6
0.4
0.2
0
0
5
10
15
20
Time s
25
30
35
FIGURE 6.8
Time response for the designs of Table 6.4: (a) }, (b) 1 , (c) .
1. From the z1 equation,
ζωn 5
1
1:5
1:5
5 1:571
5 10ln
ln
T
2cosðωd TÞ
2cosð0:5Þ In addition,
ω2d 5 ω2n ð1 2 ζ 2 Þ 5 25
Hence, we obtain the ratio
ω2d
1 2 ζ2
25
5
5
2
ζ2
ð1:571Þ2
ðζωn Þ
This gives a damping ratio ζ 5 0.3 and an undamped natural frequency ωn 5 5.24 rad/s
Finally, the z0 equation gives a gain
K 5 e22ζωn T 2 0:5 5 e2231:57130:1 2 0:5 5 0:23
2. From (6.11) and the z1 equation, we obtain
1
1
5
5 2 rad=s
τ
0:5
1
1:5eζωn T
5 10cos21 ð0:75e0:2 Þ 5 4:127 rad=s
ωd 5 cos21
2
T
ζωn 5
Solving for ζ in the equality
ω2d
1 2 ζ2
ð4:127Þ2
5
5
2
2
22
ζ
ðζωn Þ
6.2 z-Domain digital control system design
gives a damping ratio ζ 5 0.436 and an undamped natural frequency ωn 5 4.586 rad/s.
The gain for this design is
K 5 e22ζωn T 2 0:5 5 e22 3 2 3 0:1 2 0:5 5 0:17
3. For a damping ratio of 0.7, the z1 equation obtained by equating coefficients remains
nonlinear and is difficult to solve analytically. The equation now becomes
1:5 5 2cosð0:0714ωn Þe20:07ωn
The equation can be solved numerically by trial and error with a calculator to obtain the
undamped natural frequency ωn 5 3.63 rad/s. The gain for this design is
K 5 e20:14 3 3:63 2 0:5 5 0:10
This controller can also be designed graphically by drawing the root locus and a segment of the constant ζ spiral and finding their intersection. But the results obtained
graphically are often very approximate, and the solution is difficult for all but a few
simple root loci.
EXAMPLE 6.4
Consider the vehicle position control system of Example 3.3 with the transfer function
GðsÞ 5
1
sðs 1 5Þ
Design a proportional controller for the unity feedback digital control system with analog
process and a sampling period T 5 0.04 s to obtain
1. A steady-state error of 10% due to a ramp input
2. A damping ratio of 0.7
Solution
The analog transfer function together with a DAC and ADC has the z-transfer function
GZAS ðzÞ 5
7:4923 3 1024 ðz 1 0:9355Þ
ðz 2 1Þðz 2 0:8187Þ
and the closed-loop characteristic equation is
1 1 KGZAS ðzÞ 5 z2 2ð1:8187 2 7:4923 3 1024 KÞz 1 0:8187 2 7:009 3 1024 K
5 z2 2 2cosðωd TÞe2ζωn T z 1 e22ζωn T
The equation involves three parameters: ζ, ωn, and K. As in Example 6.3, equating coefficients yields two equations that we can use to evaluate two unknowns. The third parameter
must be obtained from a design specification.
1. The system is type 1, and the velocity error constant is
1 z21
KGðzÞ
Kv 5
T z
z51
7:4923 3 1024 ð1 1 0:9355Þ
5K
ð0:04Þð1 2 0:8187Þ
K
5
5
177
CHAPTER 6 Digital Control System Design
This is identical to the velocity error constant for the analog proportional control system.
In both cases, a steady-state error due to ramp of 10% is achieved with
K
100
5 Kv 5
5
eðNÞ%
100
5 10
5
10
Hence, the required gain is K 5 50.
2. As in Example 6.3, the analytical solution for constant ζ is difficult. But the design
results are easily obtained using the root locus cursor command of any CAD program.
As shown in Figure 6.9, moving the cursor to the ζ 5 0.7 contour yields a gain value of
approximately 11.7.
The root locus of Figure 6.10 shows that the critical gain for the system Kcr is approximately 268, and the system is therefore stable at the desired gain and can meet the
design specifications for both (1) and (2). However, other design criteria, such as the
damping ratio and the undamped natural frequency, should be checked. Their values can
be obtained using a CAD cursor command or by equating characteristic equation coefficients as in Example 6.3. For the gain of 50 selected in (1), the root locus of Figure 6.10
and the cursor give a damping ratio of 0.28. This corresponds to the highly oscillatory
response of Figure 6.11, which is likely to be unacceptable in practice. For the gain of
11.7 selected in (2), the steady-state error is 42.7% due to a unit ramp. It is therefore
clear that to obtain the steady-state error specified together with an acceptable transient
response, proportional control is inadequate.
1 .3
31.4
0.9
0.8
0.1
23.6
0.2
0.7
Imaginary Axis
178
0.3
0.6
15.7
0.4
0.5
0.5
0.6
0.4
System: gd
Gain: 11.7
Pole: 0.904 + 0.0889i
Damping: 0.699
Overshoot (%): 4.63
Frequency (rad/sec): 3.42
0.7
0.3
0.8
0.2
0.9
7.85
0.1
0
0
0.1
0.2
0.3
FIGURE 6.9
Root locus for the constant ζ design.
0.4
0.5
0.6
Real Axis
0.7
0.8
0.9
1
6.2 z-Domain digital control system design
1 .3
31.4
0.9
0.8
0.1
23.6
0.2
0.7
Imaginary Axis
System: gd
Gain: 268
Pole: 0.807 + 0.592i
Damping: –0.000758
Overshoot (%): 100
Frequency (rad/sec): 15.8
0.3
0.6
15.7
0.4
0.5
System: gd
Gain: 50
Pole: 0.89 + 0.246i
Damping: 0.284
Overshoot (%): 39.4
Frequency (rad/sec): 7.09
0.5
0.6
0.4
0.7
0.3
0.8
0.2
7.85
0.9
0.1
0
0.1
0
0.2
0.3
0.4
0.5
0.6
Real Axis
0.7
0.8
0.9
1
FIGURE 6.10
Root locus for K 5 50.
1.4
1.2
System: gdc
Peak amplitude: 1.4
Overshoot (%): 39.6
At time: 0.48
System: gdc
Setting Time: 2
Amplitude
1
0.8
0.6
0.4
0.2
0
0
FIGURE 6.11
Time response for K 5 50.
0.5
1
1.5
Time s
2
2.5
3
179
180
CHAPTER 6 Digital Control System Design
6.3 Digital implementation of analog controller design
This section introduces an indirect approach to digital controller design. The
approach is based on designing an analog controller for the analog subsystem and
then obtaining an equivalent digital controller and using it to digitally implement
the desired control. The digital controller can be obtained using a number of
recipes that are well known in the field of signal processing, where they are used
in the design of digital filters. In fact, a controller can be viewed as a filter that
attenuates some dynamics and accentuates others so as to obtain the desired time
response. We limit our discussion of digital filters and the comparison of various
recipes for obtaining them from analog filters to differencing methods, pole-zero
matching, and bilinear transformation. The system configuration we consider is
shown in Figure 6.12. The system includes (1) a z-transfer function model of a
DAC, analog subsystem, and ADC and (2) a cascade controller. We begin with a
general procedure to obtain a digital controller using analog design.
PROCEDURE 6.1
1. Design a controller Ca(s) for the analog subsystem to meet the desired design
specifications.
2. Map the analog controller to a digital controller C(z) using a suitable transformation.
3. Tune the gain of the transfer function C(z)GZAS(z) using proportional z-domain design
to meet the design specifications.
4. Check the sampled time response of the digital control system and repeat steps 1 to
3, if necessary, until the design specifications are met.
Step 2 of Procedure 6.1—that is, the transformation from an analog to a digital
filter—must satisfy the following requirements:
1. A stable analog filter (poles in the left half plane (LHP)) must transform to a
stable digital filter.
2. The frequency response of the digital filter must closely resemble the
frequency response of the analog filter in the frequency range 0-ωs/2 where
ωs is the sampling frequency.
Most filter transformations satisfy these two requirements to varying degrees.
However, this is not true of all analog-to-digital transformations, as illustrated by
the following section.
R(z) +
E(z)
Y(z)
U(z)
C(z)
−
FIGURE 6.12
Block diagram of a single-loop digital control system.
GZAS(z)
6.3 Digital implementation of analog controller design
6.3.1 Differencing methods
An analog filter can be represented by a transfer function or differential equation.
Numerical analysis provides standard approximations of the derivative so as to
obtain the solution to a differential equation. The approximations reduce a differential equation to a difference equation and could thus be used to obtain the difference equation of a digital filter from the differential equation of an analog
filter. We examine two approximations of the derivative: forward differencing
and backward differencing.
Forward differencing
The forward differencing approximation of the derivative is
1
½ yðk 1 1Þ 2 yðkÞ
_
yðkÞD
T
(6.13)
The approximation of the second derivative can be obtained by applying
(6.13) twice—that is,
1
_ 1 1Þ 2 yðkÞ
_ ÿðkÞD ½yðk
T
1 1
1
½ yðk 1 2Þ 2 yðk 1 1Þ 2 ½ yðk 1 1Þ 2 yðkÞ
D
T T
T
1 5 2 yðk 1 2Þ 2 2yðk 1 1Þ 1 yðkÞ
T
(6.14)
Approximations of higher-order derivatives can be similarly obtained.
Alternatively, one may consider the Laplace transform of the derivative and the
z-transform of the difference in (6.13). This yields the mapping
1
sYðsÞ- ½z 2 1YðzÞ
T
(6.15)
Therefore, the direct transformation of an s-transfer function to a z-transfer function is possible using the substitution
s-
z21
T
(6.16)
EXAMPLE 6.5: FORWARD DIFFERENCE
Apply the forward difference approximation of the derivative to the second-order analog
filter
Ca ðsÞ 5
ω2n
s2 1 2ζωn s 1 ω2n
and examine the stability of the resulting digital filter for a stable analog filter.
181
182
CHAPTER 6 Digital Control System Design
Solution
The given filter is equivalent to the differential equation
_ 1 ω2n yðtÞ 5 ω2n uðtÞ
ÿðtÞ 1 2ζωn yðtÞ
where y(t) is the filter output and u(t) is the filter input. The approximation of the first
derivative by (6.13) and the second derivative by (6.14) gives the difference equation
1
T2
yðk 1 2Þ 2 2yðk 1 1Þ 1 yðkÞg 1 2ζωn
1
½yðk 1 1Þ 2 yðkÞ 1 ω2n yðkÞ 5 ω2n uðkÞ
T
Multiplying by T2 and rearranging terms, we obtain the digital filter
yðk 1 2Þ 1 2½ζωn T 2 1yðk 1 1Þ 1 ½ðωn TÞ2 2 2ζωn T 1 1yðkÞ 5 ðωn TÞ2 uðkÞ
Equivalently, we obtain the transfer function of the filter using the simpler transformation
(6.16)
ω2n
CðzÞ 5 2
s 12ζωn s1ω2n z21
s5 T
5
z2
ðωn TÞ2
1 2½ζωn T 2 1z 1 ½ðωn TÞ2 2 2ζωn T 1 1
For a stable analog filter, we have ζ . 0 and ωn . 0 (positive denominator coefficients
are sufficient for a second-order polynomial). However, the digital filter is unstable if the
magnitude of the constant term in its denominator polynomial is greater than unity. This
gives the instability condition
ðωn TÞ2 2 2ζωn T 1 1 . 1
i:e:; ζ , ωn T=2
For example, a sampling period of 0.2 s and an undamped natural frequency of
10 rad/s yield unstable filters for any underdamped analog filter.
Backward differencing
The backward differencing approximation of the derivative is
1
½ yðkÞ 2 yðk 2 1Þ
_
yðkÞD
T
(6.17)
The approximation of the second derivative can be obtained by applying
(6.17) twice—that is,
1
_ 2 yðk
_ 2 1Þ
ÿðkÞD ½yðkÞ
T
1 1
1
½yðkÞ 2 yðk 2 1Þ 2 ½ yðk 2 1Þ 2 yðk 2 2Þ
D
T T
T
1 5 2 yðkÞ 2 2yðk 2 1Þ 1 yðk 2 2Þ
T
(6.18)
6.3 Digital implementation of analog controller design
Approximations of higher-order derivatives can be similarly obtained. One
may also consider the Laplace transform of the derivative and the z-transform of
the difference in (6.17). This yields the substitution
s-
z21
zT
(6.19)
EXAMPLE 6.6: BACKWARD DIFFERENCE
Apply the backward difference approximation of the derivative to the second-order analog
filter
Ca ðsÞ 5
s2
ω2n
1 2ζωn s 1 ω2n
and examine the stability of the resulting digital filter for a stable analog filter.
Solution
We obtain the transfer function of the filter using (6.19)
ω2n
CðzÞ 5 2
2
s 12ζωn s1ωn z21
s5 zT
5
ðωn TzÞ2
½ðωn TÞ 1 2ζωn T 1 1z2 2 2½ζωn T 1 1z 1 1
2
The stability conditions for the digital filter are (see Chapter 4)
½ðωn TÞ2 1 2ζωn T 1 1 1 2½ζωn T 1 1 1 1 . 0
½ðωn TÞ2 1 2ζωn T 1 1 2 1 . 0
½ðωn TÞ2 1 2ζωn T 1 1 2 2½ζωn T 1 1 1 1 . 0
The conditions are all satisfied for ζ . 0 and ωn . 0—that is, for all stable analog filters.
6.3.2 Pole-zero matching
We know from equation (6.3) that discretization maps an s-plane pole at ps to a
z-plane pole at epsT but that no rule exists for mapping zeros. In pole-zero matching, a discrete approximation is obtained from an analog filter by mapping both
poles and zeros using (6.3). If the analog filter has n poles and m zeros, then we
say that the filter has n 2 m zeros at infinity. For n 2 m zeros at infinity, we add
n 2 m or n 2 m 2 1 digital filter zeros at unity. If the zeros are not added, it can
be shown that the resulting system will include a time delay (see Problem 6.5).
The second choice gives a strictly proper filter where the computation of the output is easier, since it only requires values of the input at past sampling points.
Finally, we adjust the gain of the digital filter so that it is equal to that of the analog filter at a critical frequency dependent on the filter. For a low-pass filter, α is
selected so that the gains are equal at DC; for a bandpass filter, they are set equal
at the center of the pass band.
183
184
CHAPTER 6 Digital Control System Design
For an analog filter with transfer function
m
Ga ðsÞ 5 K
Li51 ðs 2 ai Þ
m
Lj51 ðs 2 bj Þ
(6.20)
we have the digital filter
m
GðzÞ 5 αK
ðz11Þn2m21 Li51 ðz 2 eai T Þ
m
Lj51 ðz 2 ebj T Þ
(6.21)
where α is a constant selected for equal filter gains at a critical frequency. For
example, for a low-pass filter, α is selected to match the DC gains using Ga(1) 5
Ga(0), while for a high-pass filter, it is selected to match the high-frequency
gains using G(21) 5 Ga(N). Setting z 5 e ( jωT) 5 21 (i.e., ωT = π) is equivalent
to selecting the folding frequency ωs /2, which is the highest frequency allowable
without aliasing. Pole-zero matched digital filters can be obtained using the
MATLAB command
.. g 5 c2d ðga; T; 'matched'Þ
EXAMPLE 6.7
Find a pole-zero matched digital filter approximation for the analog filter
Ga ðsÞ 5
ω2n
s2 1 2ζωn s 1 ω2n
If the damping ratio is equal to 0.5 and the undamped natural frequency is 5 rad/s, determine the transfer function of the digital filter for a sampling period of 0.1 s. Check your
answer using MATLAB and obtain the frequency response of the digital filter.
Solution
The filter has a zero at the origin and complex conjugate poles at s1,2 5 2ζωn 6 jωd. We
apply the pole-zero matching transformation to obtain
GðzÞ 5
αðz 1 1Þ
z2 2 2e2ζωn T cosðωd TÞz 1 e22ζωn T
The analog filter has two zeros at infinity, and we choose to add one digital filter zeros
at 21 for a strictly proper filter. The difference equation for the filter is
yðk 1 2Þ 5 2e2ζωn T cosðωd TÞyðk 1 1Þ 2 e22ζωn T yðkÞ
1αðuðk 1 1Þ 2 uðkÞÞ
Thus, the computation of the output only requires values of the input at earlier sampling
points, and the filter is easily implementable. The gain α is selected in order for the digital
filter to have the same DC gain of the analog filter, which is equal to unity.
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
For the given numerical values, we have the damping ratio ωd 5 5 1 2 0:52 5
4:33 rad=s and the filter transfer function
GðzÞ 5
0:09634ðz 1 1Þ
z2 2 1:414z 1 0:6065
6.3 Digital implementation of analog controller design
The following MATLAB commands give the transfer function:
.. wn 5 5;zeta 5 0.5; % Undamped natural frequency, damping ratio
.. ga 5 tf([wn^2],[1,2 zeta wn,wn^2]) % Analog transfer function
Transfer function:
25
–––––––––
s^2 1 5 s 1 25
.. g 5 c2d(ga,.1,'matched') % Transformation with a sampling period 0.1
Transfer function:
0.09634 z 1 0.09634
–––––––––––
z^2 2 1.414 z 1 0.6065
Sampling time: 0.1
The frequency responses of the analog and digital filters obtained using
MATLAB are shown in Figure 6.13. Note that the frequency responses are
almost identical in the low frequency range but become different at high
frequencies.
Bode Diagram
10
Magnitude (dB)
0
–10
Ga(s)
–20
–30
–40
G(z)
–50
–60
–70
–80
0
Phase (deg)
–45
–90
–135
–180
–225
–270
10–1
100
101
Frequency (rad/sec)
FIGURE 6.13
Frequency response of digital filter for Example 6.7.
102
185
186
CHAPTER 6 Digital Control System Design
6.3.3 Bilinear transformation
The relationship
s5c
z21
z11
(6.22)
with a linear numerator and a linear denominator and a constant scale factor c is
known as a bilinear transformation. The relationship can be obtained from the
equality z 5 esT using the first-order approximation
s5
1
2 z21
lnðzÞD
T
T z11
(6.23)
where the constant c 5 2/T. A digital filter C(z) is obtained from an analog filter
Ca(s) by the substitution
(6.24)
CðzÞ5Ca ðsÞs5c½z21
z11
The resulting digital filter has the frequency response
jωT
Cðe Þ 5 Ca ðsÞ
ejωT 21
s5c
5 Ca c
ejωT 11
ejωT=2 2 e2jωT=2
ejωT=2 1 e2jωT=2
Thus, the frequency responses of the digital and analog filters are related by
Cðe
jωT
Þ 5 Ca
ωT
jctan
2
(6.25)
Evaluating the frequency response at the folding frequency ωs/2 gives
0
Cðejωs T=2 Þ
2
31
ωs T 5A
5 Ca @jctan4
4
0
2 31
2π
5 Ca @jctan4 5A 5 Ca ðjNÞ
4
We observe that bilinear mapping squeezes the entire frequency response of the
analog filter for a frequency range 0-N into the frequency range 0-ωs/2. This
implies the absence of aliasing (which makes the bilinear transformation a popular method for digital filter design) but also results in distortion or warping of the
6.3 Digital implementation of analog controller design
frequency response. The relationship between the frequency ωa of the analog filter
and the associated frequency ω of the digital filter for the case c 5 2/T—namely,
2
ωT
ωa 5 tan
T
2
is plotted in Figure 6.14 for T 5 1. Note that, in general, if the sampling period is
sufficiently small so that ω{π/T, then
ωT
ωT
tan
D
2
2
and therefore ωa ω, so that the effect of the warping is negligible.
In any case, the distortion of the frequency response can be corrected at a single frequency ω0 using the prewarping equality
ω0 T
jω0 T
5 Ca ðjω0 Þ
(6.26)
Cðe Þ 5 Ca jctan
2
The equality holds provided that the constant c is chosen as
c5
ω0
tan ω20 T
(6.27)
The choice of the prewarping frequency ω0 depends on the mapped filter.
In control applications, a suitable choice of ω0 is the 3-dB frequency for a PI or
PD controller and the upper 3-dB frequency for a PID controller. This is explored
further in design examples.
Digital Filter Frequency
3
2.5
2
1.5
1
0.5
0
0
2
4
6
8
10 12 14
Analog Filter Frequency
16
18
20
FIGURE 6.14
Relationship between analog filter frequencies ωa and the associated digital filter
frequencies with bilinear transformation.
187
CHAPTER 6 Digital Control System Design
In MATLAB, the bilinear transformation is accomplished using the following
command:
.. gd 5 c2dðg; tc; 'tustin'Þ
where g is the analog system and tc is the sampling period. If prewarping is
requested at a frequency w, then the command is
.. gd 5 c2dðg; tc; 'prewarp'; wÞ
EXAMPLE 6.8
Design a digital filter by applying the bilinear transformation to the analog filter
Ca ðsÞ 5
1
0:1s 1 1
(6.28)
with T 5 0.1 s. Examine the warping effect and then apply prewarping at the 3-dB
frequency.
Solution
By applying the bilinear transformation (6.22) to (6.28), we obtain
CðzÞ 5
1
z11
5
2 z21
3z 2 1
0:1 0:1
z11 1 1
The Bode plots of Ca(s) (solid line) and C(z) (dash-dot line) are shown in Figure 6.15,
where the warping effect can be evaluated. We select the 3-dB frequency ω0 5 10 as a prewarping frequency and apply (6.27) to obtain
CðzÞ 5
1
0:35z 1 0:35
D
z21
z 2 0:29
0:1 tan 10
1
1
z11
ð10U0:1
2 Þ
0
–1
Magnitude (dB)
188
–2
–3
–4
–5
–6 0
10
101
Frequency (rad/sec)
FIGURE 6.15
Bode plots of the analog filter (solid) and the digital filter obtained with (dashed) and
without prewarping (dash-dot).
6.3 Digital implementation of analog controller design
The corresponding Bode plot is shown again in Figure 6.15 (dashed line). It coincides
with the Bode plot of C(s) at ω0 5 10. Note that for lower values of the sampling period,
the three Bode plots tend to coincide.
Another advantage of bilinear transformation is that it maps points in the LHP
to points inside the unit circle and thus guarantees the stability of a digital filter
for a stable analog filter. This property was discussed in Section 4.42 and is clearly
demonstrated in Figure 4.4.
Bilinear transformation of the analog PI controller gives the following digital
PI controller:
ðs1aÞ CðzÞ 5 K
s z21
s5c½z11
(6.29)
a2c
a 1 c z 1 a1c
5K
c
z21
The digital PI controller increases the type of the system by one and can therefore be used to improve steady-state error. As in the analog case, it has a zero
that reduces the deterioration of the transient response due to the increase in system type. The PI controller of (6.29) has a numerator order equal to its denominator order. Hence, the calculation of its output from its difference equation requires
knowledge of the input at the current time. Assuming negligible computational
time, the controller is approximately realizable.
Bilinear transformation of the analog PD controller gives the digital PD
controller
CðzÞ 5 Kðs1aÞ
z21
s5c½z11
(6.30)
5 Kða 1 cÞ
c
z 1 aa 2
1c
z11
This includes a zero that can be used to improve the transient response and a
pole at z 5 21 that occurs because the continuous time system is not proper (see
Problem 6.8). A pole at z 5 21 corresponds to an unbounded frequency response
at the folding frequency, as ejωs T=2 5 ejπ 5 21, and must therefore be eliminated.
However, eliminating the undesirable pole would result in an unrealizable controller. An approximately realizable PD controller is obtained by replacing the pole
at z 5 21 with a pole at the origin to obtain
CðzÞ 5 Kða 1 cÞ
z1
a2c
a1c
z
(6.31)
189
190
CHAPTER 6 Digital Control System Design
A pole at the origin is associated with a term that decays as rapidly as possible
so as to have the least effect on the controller dynamics. However, this variation
from direct transformation results in additional distortion of the analog filter and
complication of the digital controller design and doubles the DC gain of the controller. To provide the best approximation of the continuous-time controller, disregarding subsequent gain tuning, the gain K can be halved.
Bilinear transformation of the analog PID controller gives the digital PD
controller
ðs1aÞðs1bÞ CðzÞ 5 K
z21
s
s5c½z11
5K
c
c
z 1 bb 2
ða 1 cÞðb 1 cÞ z 1 aa 2
1c
1c
ðz 1 1Þðz 2 1Þ
c
The controller has two zeros that can be used to improve the transient
response and a pole at z 5 1 to improve the steady-state error. As with PD control,
transforming an improper transfer function yields a pole at z 5 21, which must
be replaced by a pole at the origin to yield a transfer function with a bounded frequency response at the folding frequency. The resulting transfer function is
approximately realizable and is given by
CðzÞ 5 K
ða 1 cÞðb 1 cÞ z 1
c
a2c
a1c
z1
zðz 2 1Þ
b2c
b1c
(6.32)
As in the case of PD control, the modification of the bilinearly transformed
transfer function results in distortion that can be reduced by halving the
gain K.
Using Procedure 6.1 and equations (6.29), (6.31), and (6.32), respectively, digital PI, PD, and PID controllers can be designed to yield satisfactory transient and
steady-state performance.
EXAMPLE 6.9
Design a digital controller for a DC motor speed control system (see Example 3.6) where
the (type 0) analog plant has the transfer function
GðsÞ 5
1
ðs 1 1Þðs 1 10Þ
to obtain zero steady-state error due to a unit step, a damping ratio of 0.7, and a settling
time of about 1 s.
Solution
The design is completed following Procedure 6.1. First, an analog controller is designed
for the given plant. For zero steady-state error due to unit step, the system type must be
increased by one. A PI controller affects this increase, but the location of its zero must be
6.3 Digital implementation of analog controller design
chosen so as to obtain an acceptable transient response. The simplest possible design is
obtained by pole-zero cancellation and is of the form
Ca ðsÞ 5 K
s11
s
The corresponding loop gain is
Ca ðsÞGðsÞ 5
K
sðs 1 10Þ
Hence, the closed-loop characteristic equation of the system is
sðs 1 10Þ 1 K 5 s2 1 2ζωn s 1 ω2n
Equating coefficients gives ζωn 5 5 rad/s and the settling time
Ts 5
4
4
5 5 0:8 s
ζωn
5
as required. The damping ratio of the analog system can be set equal to 0.7 by appropriate
choice of the gain K. The gain selected at this stage must often be tuned after filter transformation to obtain the same damping ratio for the digital controller. We solve for the
undamped natural frequency
ωn 5 10=ð2ζÞ 5 10=ð2 3 0:7Þ 5 7:142 rad=s
The corresponding analog gain is
K 5 ω2n 5 51:02
1 .5
62.8
0.9
0.8
0.1
Imaginary Axis
47.1
0.2
0.7
0.3
0.6
31.4
0.4
0.5
0.5
0.4
0.6
0.3
0.7
System: l
Gain: 46.7
Pole: 0.904 + 0.0888i
Damping: 0.699
Overshoot (%): 4.63
Frequency (rad/sec): 6.85
0.8
0.2
0.9
0.1
0
0
0.1
0.2
0.3
0.4
0.5
0.6
Real Axis
FIGURE 6.16
Root locus for PI design.
0.7
0.8
15.7
0.9
1
191
CHAPTER 6 Digital Control System Design
We therefore have the analog filter
Ca ðsÞ 5 51:02
s11
s
Next, we select a suitable sampling period for an undamped natural frequency of about
7.14 rad/s. We select T 5 0.02 s , 2π/(40ωd), which corresponds to a sampling frequency
higher than 40 times the damped natural frequency (see Chapter 2). The model of the
analog plant together with an ADC and sampler is
GðsÞ
GZAS ðzÞ 5 ð1 2 z21 Þ Z
s
z 1 0:9293
24
5 1:8604 3 10
ðz 2 0:8187Þðz 2 0:9802Þ
Bilinear transformation of the PI controller, with gain K included as a free parameter, gives
CðzÞ 5 1:01K
z 2 0:9802
z21
Because the analog controller was obtained using pole-zero cancellation, near polezero cancellation occurs when the digital controller C(z) is multiplied by GZAS(z). The gain
can now be tuned for a damping ratio of 0.7 using a CAD package with the root locus of
the loop gain C(z)GZAS(z). From the root locus, shown in Figure 6.16, at ζ 5 0.7, the gain K
is about 46.7, excluding the 1.01 gain of C(z) (i.e., a net gain of 47.2). The undamped
natural frequency is ωn 5 6.85 rad/s. This yields the approximate settling time
Ts 5
4
4
5
ζωn
6:85 3 0:7
5 0:83 s
1.4
System: gcl
Peak amplitude: 10.5
Overshoot (%): 4.74
At time: 0.64
1.2
System: gcl
Setting Time: 0.88
1
Amplitude
192
0.8
0.6
0.4
0.2
0
0
0.2
0.4
FIGURE 6.17
Step response for PI design with K 5 47.2.
0.6
Time s
0.8
1
1.2
6.3 Digital implementation of analog controller design
The settling time is acceptable but is slightly worse than the settling time for the analog controller. The step response of the closed-loop digital control system shown in
Figure 6.17 is also acceptable and confirms the estimated settling time. The net gain
value of 47.2, which meets the design specifications, is significantly less than the gain
value of 51.02 for the analog design. This demonstrates the need for tuning the controller
gain after mapping the analog controller to a digital controller.
EXAMPLE 6.10
Design a digital controller for the DC motor position control system of Example 3.6, where
the (type 1) analog plant has the transfer function
GðsÞ 5
1
sðs 1 1Þðs 1 10Þ
to obtain a settling time of about 1 second and a damping ratio of 0.7.
Solution
Using Procedure 6.1, we first observe that an analog PD controller is needed to improve
the system transient response. Pole-zero cancellation yields the simple design
Ca ðsÞ 5 Kðs 1 1Þ
We can solve for the undamped natural frequency analytically, or we can use a CAD
package to obtain the values K 5 51.02 and ωn 5 7.143 rad/s for ζ 5 0.7.
System: l
Gain: 2.16e + 0.03
Pole: 0.909 + 0.0854i
Damping: 0.695
Overshoot (%): 4.8
Frequency (rad/sec): 6.51
0.2
Imaginary Axis
0.1
0
–0.1
–0.2
–0.3
0.4
FIGURE 6.18
Root locus for PD design.
0.5
0.6
0.7
0.8 0.9
1
Real Axis
1.1
1.2
1.3
1.4
193
CHAPTER 6 Digital Control System Design
1.4
System: cl
Peak amplitude: 1.05
Overshoot (%): 4.76
At time: 0.68
1.2
System: cl
Setting Time: 0.94
1
Amplitude
194
0.8
0.6
0.4
0.2
0
0
0.5
1
1.5
Time s
FIGURE 6.19
Time step response for PD design with K 5 2,160.
A sampling period of 0.02 s is appropriate because it is less than 2π/(40ωd). The plant
with the ADC and DAC has the z-transfer function
(
)
GðsÞ
ðz 1 0:2534Þðz 1 3:535Þ
GZAS ðzÞ 5 ð1 2 z21 Þ Z
5 1:2629 3 1026
s
ðz 2 1Þðz 2 0:8187Þðz 2 0:9802Þ
Bilinear transformation of the PD controller gives
CðzÞ 5 K
z 2 0:9802
5 Kð1 2 0:9802z21 Þ
z
The root locus of the system with PD control (Figure 6.18) gives a gain K of 2,160
and an undamped natural frequency of 6.51 rad/s at a damping ratio ζ 5 0.7. The settling
time for this design is about
Ts 5
4
4
5
5 0:88 s
ζωn
0:7 3 6:51
which meets the design specifications.
Checking the step response with MATLAB gives Figure 6.19 with a settling time of
0.94 s, a peak time of 0.68 s, and a 5% overshoot. The time response shows a slight
deterioration from the characteristics of the analog system but meets all the design
specifications. In some cases, the deterioration may necessitate repeatedly modifying the
digital design or modifying the analog design and then mapping it to the z-domain until
the resulting digital filter meets the desired specifications.
Note that for a prewarping frequency ω0 5 1 rad/s, the 3-dB frequency of the PD controller, ω0T 5 0.02 rad and tan(ω0T/2) 5 tan(0.01)D0.01. Hence, equation (6.25) is approximately valid without prewarping, and prewarping has a negligible effect on the design.
6.3 Digital implementation of analog controller design
EXAMPLE 6.11
Design a digital controller for a speed control system, where the analog plant has transfer
function
GðsÞ 5
1
ðs 1 1Þðs 1 3Þ
to obtain a time constant of less than 0.3 second, a dominant pole damping ratio of at
least 0.7, and zero steady-state error due to a step input.
Solution
The root locus of the analog system is shown in Figure 6.20. To obtain zero steady-state
error due to a step input, the system type must be increased to one by adding an integrator
in the forward path. However, adding an integrator results in significant deterioration of
the time response or in instability. If the pole at 21 is canceled, the resulting system is
stable but has ζωn 5 1.5—that is, a time constant of 2/3 s and not less than 0.3 s as specified. Using a PID controller provides an additional zero that can be used to stabilize the
system and satisfy the remaining design requirements.
For a time constant τ of 0.3 s, we have ζωn 5 1/τ $ 3.33 rad/s. A choice of ζ 5 0.7 and
ωn of about 6 rad/s meets the design specifications. The design appears conservative, but
we choose a larger undamped natural frequency than the minimum needed in anticipation
of the deterioration due to adding PI control. We first design a PD controller to meet these
specifications using MATLAB. We obtain the controller angle of about 52.4 using the
angle condition. The corresponding zero location is
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
6 1 2 ð0:7Þ2
1ð0:7Þð6ÞD7:5
a5
tanð52:4 Þ
The root locus for the system with PD control (Figure 6.21) shows that the system with
ζ 5 0.7 has ωn of about 6 rad/s and meets the transient response specifications with a gain
j
5
n=
6
4
3
2
ζ = 0.7
1
0
–1
–5
–4
–3
FIGURE 6.20
Root locus for the analog speed control system.
–2
–1
0 σ
195
196
CHAPTER 6 Digital Control System Design
jω
6
5
ωn = 6
4
3
ζ= 0.7
2
1
0
–6
–5
–4
–3
–2
–1
0 σ
FIGURE 6.21
Root locus of a PD-controlled system.
of 4.4 and ζωn 5 4.2. Following the PI-design procedure, we place the second zero of the
PID controller at one-tenth this distance from the jω axis to obtain
Ca ðsÞ 5 K
ðs 1 0:4Þðs 1 7:5Þ
s
To complete the analog PID design, the gain must be tuned to ensure that ζ 5 0.7.
Although this step is not needed, we determine the gain K 5.8, and ωn 5 6.7 rad/s
(Figure 6.22) for later comparison to the actual gain value used in the digital design. The
analog design meets the transient response specification with ζωn 5 4.69 . 3.33, and the
dynamics allow us to choose a sampling period of 0.025 s (ωs . 50ωd).
The model of the analog plant with DAC and ADC is
GðsÞ
GZAS ðzÞ 5 ð1 2 z21 ÞZ
s
z 1 0:936
23
5 1:170 3 10
ðz 2 0:861Þðz 2 0:951Þ
Bilinear transformation and elimination of the pole at 2 1 yields the digital PID
controller
CðzÞ 5 47:975K
ðz 2 0:684Þðz 2 0:980Þ
zðz 2 1Þ
The root locus for the system with digital PID control is shown in Figure 6.23, and the
system is seen to be minimum phase (i.e., its zeros are inside the unit circle).
For design purposes, we zoom in on the most significant portion of the root locus and
obtain the plot of Figure 6.24. With K 5 9.62, 13.2, the system has ζ 5 0.7, 0.5, and ωn 5 43.2,
46.2 rad/s, respectively. Both designs have a sufficiently fast time constant, but the second
damping ratio is less than the specified value of 0.7. The time response of the two digital systems and for analog control with K 5 100 are shown in Figure 6.25. Lower gains give an unacceptably slow analog design. The time response for the high-gain digital design is very fast.
However, it has an overshoot of over 4% but has a settling time of 5.63 s. The digital design
for ζ 5 0.7 has a much slower time response than its analog counterpart.
6.3 Digital implementation of analog controller design
jω
6
System: 1
Gain: 5.76
Pole: –4.69 + 4.78i
Damping: 0.70
Overshoot (%): 4.57
Frequency (rad/sec): 6.7
5
4
ωn = 6
3
2
ζ = 0.7
1
0
–6
–5
–4
–3
–2
–1
0 σ
FIGURE 6.22
Root locus of an analog system with PID control.
1.5
Imaginary Axis
1
0.5
0
–0.5
–1
–1.5
–2.5
–2
–1.5
–1
–0.5
Real Axis
0
0.5
1
FIGURE 6.23
Root locus of a system with digital PID control.
It is possible to improve the design by trial and error, including redesign of the analog
controller, but the design with ζ 5 0.5 may be acceptable. One must weigh the cost of redesign against that of relaxing the design specifications for the particular application at
hand. The final design must be a compromise between speed of response and relative
stability.
197
CHAPTER 6 Digital Control System Design
1 π/T
0.9
0.4π/T
0.8
0.1
Imaginary Axis
0.6
0.5
0.2π/T
0.5
0.4
0.6
0.7
0.3
System: l
0.8
Gain: 9.65
Pole: 0.337 + 0.327i
0.9
Damping: 0.701
Overshoot (%): 4.57
Frequency (rad/sec): 43.2
0.2
0.1
0
0.3π/T
System: l
0.213.2
Gain:
Pole: 0.303 + 0.471i
0.3
Damping:
0.502
Overshoot (%): 16.2
0.4
Frequency (rad/sec): 46.2
0.7
0
0.1
0.2
0.3
0.1π/T
0.4
0.5
0.6
Real Axis
0.7
0.8
0.9
1
FIGURE 6.24
Detail of the root locus of a system with digital PID control.
Digital 13.2
Analog 100
1
Digital 9.62
0.8
Amplitude
198
0.6
0.4
0.2
0
0
1
2
3
4
Time s
5
6
7
8
FIGURE 6.25
Time step response for the digital PID design with K 5 9.62 (light gray), K 5 13.2 (dark
gray), and for analog design (black).
6.3 Digital implementation of analog controller design
6.3.4 Empirical digital PID controller tuning
As explained in Section 5.5, the parameters of a PID controller are often selected
by means of tuning rules. This concept can also be exploited to design a digital
PID controller. The reader can show (Problem 6.7) that bilinear transformation of
the PID controller expression (5.20) yields
1 T z11
2 z21
1 Td
(6.33)
CðzÞ 5 Kp 1 1
Ti 2 z 2 1
T z11
Bilinear transformation of the PID controller results in a pole at z 5 21
because the derivative part is not proper (see Section 12.4.1). As in Section 6.3.3,
we avoid an unbounded frequency response at the folding frequency by replacing
the pole at z 5 21 with a pole at z 5 0 and dividing the gain by two. The resulting
transfer function is
Kp
T z 1 1 2Td z 2 1
11
1
CðzÞ 5
2
T
2Ti z 2 1
z
If parameters Kp, Ti, and Td are obtained by means of a tuning rule as in the
analog case, then the expression of the digital controller is obtained by substituting in the previous expression. The transfer function of a zero-order hold can be
approximated by truncating the series expansions as
GZOH ðsÞ 5
1 2 e2sT 1 2 1 1 Ts 2 ðTsÞ2 =2 1 ?
Ts
T
D
512
1 ?De22s
s
Ts
2
Thus, the presence of the ZOH can be considered as an additional time delay
equal to half of the sampling period. The tuning rules of Table 5.1 can then be
applied to a system with a delay equal to the sum of the process time delay and a
delay of T/2 due to the zero-order hold.
EXAMPLE 6.12
Design a digital PID controller with sampling period T 5 0.1 for the analog plant of Example 5.9
GðsÞ 5
1
e20:2s
ðs11Þ4
by applying the Ziegler-Nichols tuning rules of Table 5.1.
Solution
A first-order-plus-dead-time model of the plant was obtained in Example 5.9 using the tangent method with gain cK 5 1, a dominant time constant t 5 3, and an apparent time delay
L 5 1.55. The apparent time delay for digital control is obtained by adding half of the value
of the sampling period (0.05). This gives L 5 1.55 1 0.05 5 1.6. The application of the tuning rules of Table 5.1 yields
τ
5 2:25
Kp 5 1:2
KL
Ti 5 2L 5 3:2
Td 5 0:5L 5 0:8
199
CHAPTER 6 Digital Control System Design
1.5
1
Output
200
0.5
0
0
10
20
30
40
50 60
Time s
70
80
90 100
FIGURE 6.26
Process output with the digital PID controller tuned with the Ziegler-Nichols method.
Thus, the digital PID controller has the transfer function
CðzÞ 5
19:145z2 2 35:965z 1 16:895
zðz 2 1Þ
The response of the digital control system due to a unit step reference input applied
at time t 5 0 and to a unit step change in the control variable at time t 5 50 is shown in
Figure 6.26. The response is similar to the result obtained with the analog PID controller
in Example 5.9.
6.4 Direct z-domain digital controller design
Obtaining digital controllers from analog designs involves approximation that
may result in significant controller distortion. In addition, the locations of the controller poles and zeros are often restricted to subsets of the unit circle. For example, bilinear transformation of the term (s 1 a) gives [z 2 (c 2 a)/(c 1 a)], as seen
from (6.29) through (6.32). This yields only RHP zeros because a is almost
always smaller than c. The plant poles are governed by pz 5 eps T , where ps and pz
are the s-domain and z-domain poles, respectively, and can be canceled with RHP
zeros. Nevertheless, the restrictions on the poles and zeros in (6.29) through
(6.32) limit the designer’s ability to reshape the system root locus.
Another complication in digital approximation of analog filters is the need to
have a pole at 0 in place of the pole at 21, as obtained by direct digital transformation, to avoid an unbounded frequency response at the folding frequency. This
may result in a significant difference between the digital and analog controllers
and may complicate the design process considerably.
6.4 Direct z-domain digital controller design
Alternatively, it is possible to directly design controllers in the less familiar zplane. The controllers used are typically of the same form as those discussed in
Section 6.3, but the poles of the controllers are no longer restricted as in sdomaintoz-domain mapping. Thus, the poles can now be in either the LHP or
the RHP as needed for design.
Because of the similarity of s-domain and z-domain root loci, Procedures 5.1,
5.2, and 5.3 are applicable with minor changes in the z-domain. However, equation (5.14) is no longer valid because the real and imaginary components of the
complex conjugate poles are different in the z-domain. In addition, a digital controller with a zero and no poles is not realizable, and a pole must be added to the
controller. To minimize the effect of the pole on the time response, it is placed at
the origin. This is analogous to placing an s-plane pole far in the LHP to minimize its effect. We must now derive an expression similar to (5.14) for the digital
PID controller.
Using the expression for the complex conjugate z-domain poles (6.4), we
obtain (Figure 6.27)
a 5 Refzcl g 2
2ζωn T
5e
Imfzcl g
tanðθa Þ
e2ζωn T sinðωd TÞ
cosðωd TÞ 2
tanðθa Þ
(6.34)
where θa is the angle of the controller zero. θa is given by
θa 5 θc 1 θp
5 θc 1 θzcl
(6.35)
In (6.35), θzcl is the angle of the controller pole and θc is the controller angle
contribution at the closed-loop pole location. The sign of the second term in
(6.34) is negative, unlike (5.14), because the real part of a stable z-domain pole
can be positive. In addition, a digital controller with a zero and no poles is not
realizable and a pole, at the origin or other locations inside the unit circle, must
Im{z}
zcl
Im{zcl}
θc
a
Re{zcl}
FIGURE 6.27
PD compensator zero.
Re{z}
201
202
CHAPTER 6 Digital Control System Design
first be added to the system before computing the controller angle. The computation of the zero location is simple using the following MATLAB function:
% Digital PD controller design with pole at origin and zero to be
selected.
function[c,zcl] 5 dpdcon(zeta,wn,g,T)
% g (l) is the uncompensated (compensated) loop gain
% zeta and wn specify the desired closed-loop pole, T 5 sampling period
% zcl is the closed-loop pole, theta is the angle of the
% compensator zero at zcl.
% The corresponding gain is “k” and the zero is “a”.
wdT 5 wn T sqrt(1-zeta^2); % Pole angle: T damped natural frequency
rzcl 5 exp(-zeta wn T) cos(wdT); % Real part of the closed-loop pole.
izcl 5 exp(-zeta wn T) sin(wdT); % Imaginary part of the closed-loop
pole.
zcl 5 rzcl 1 j izcl; % Complex closed-loop pole.
% Find the angle of the compensator zero. Include the contribution
% of the pole at the origin.
theta 5 pi - angle(evalfr(g,zcl)) 1 angle(zcl);
a 5 rzcl-izcl/tan(theta); % Calculate the zero location.
c 5 zpk([a],[0],1,T); % Calculate the compensator transfer function.
L 5 c g; % Loop gain.
k 5 1/abs(evalfr( l, zcl)); % Calculate the gain.
c 5 k c; %Include the correct gain.
Although equation (6.34) is useful in some situations, in many others, PD or
PID controllers can be more conveniently obtained by simply canceling the slow
poles of the system with zeros. Design by pole-zero cancellation is the simplest
possible and should be explored before more complex designs are attempted.
The main source of difficulty in z-domain design is the fact that the stable region
is now the unit circle as opposed to the much larger left half of the s-plane. In addition, the selection of pole locations in the z-domain directly is less intuitive and is
generally more difficult than s-domain pole selection. Pole selection and the entire
design process are significantly simplified by the use of CAD tools and the availability of constant ζ contours, constant ωn contours, and cursor commands.
No new theory is needed to introduce z-domain design, and we proceed
directly to design examples. We repeat Examples 6.9, 6.10, and 6.11 using direct
digital design to demonstrate its strengths and weaknesses compared to the indirect design approach.
EXAMPLE 6.13
Design a digital controller for a DC motor speed control system where the (type 0) analog
plant has the transfer function
GðsÞ 5
1
ðs 1 1Þðs 1 10Þ
6.4 Direct z-domain digital controller design
to obtain zero steady-state error due to a unit step, a damping ratio of 0.7, and a settling
time of about 1 s.
Solution
First, selecting T 5 0.1 s, we obtain the z-transfer function
GðsÞ
z 1 0:694
5 3:55 3 1023
GZAS ðzÞ 5 ð1 2 z21 ÞZ
s
ðz 2 0:368Þðz 2 0:905Þ
To perfectly track a step input, the system type must be increased by one and a PI
controller is therefore used. The controller has a pole at z 5 1 and a zero to be selected to
meet the remaining design specifications. Directly canceling the pole at z 5 0.905 gives a
design that is almost identical to that of Example 6.9 and meets the design specifications.
EXAMPLE 6.14
Design a digital controller for the DC motor position control system of Example 3.6, where
the (type 1) analog plant has the transfer function
GðsÞ 5
1
sðs 1 1Þðs 1 10Þ
for a settling time of less than 1 s and a damping ratio of 0.7.
Solution
For a sampling period of 0.01 s, the plant, ADC, and DAC have the z-transfer function
GðsÞ
GZAS ðzÞ 5 ð1 2 z21 ÞZ
s
ðz 1 0:2606Þðz 1 3:632Þ
27
5 1:6217 3 10
ðz 2 1Þðz 2 0:9048Þðz 2 0:99Þ
Using a digital PD controller improves the system transient response as in Example
6.10. Pole-zero cancellation yields the simple design
CðzÞ 5 K
z 2 0:99
z
which includes a pole at z 5 0 to make the controller realizable. The design is almost identical to that of Example 6.10 and meets the desired transient response specifications with
a gain of 4,580.
EXAMPLE 6.15
Design a digital controller for the DC motor speed control system where the analog plant
has the transfer function
GðsÞ 5
1
ðs 1 1Þðs 1 3Þ
for a time constant of less than 0.3 second, a dominant pole damping ratio of at least 0.7,
and zero steady-state error due to a step input.
203
CHAPTER 6 Digital Control System Design
2
1.5
1
Imaginary Axis
204
0.5
0
–0.5
–1
–1.5
–2
–5
–4
–3
–2
Real Axis
–1
0
1
FIGURE 6.28
Root locus for digital PI control.
Solution
The plant is type 0 and is the same as in Example 6.10 with a sampling period
T 5 0.005 s:
GðsÞ
z 1 0:9934
5 1:2417 3 1025
GZAS ðzÞ 5 ð1 2 z21 ÞZ
s
ðz 2 0:9851Þðz 2 0:995Þ
For zero steady-state error due to step, the system type must be increased by one by adding a pole at z 5 1. A controller zero can be used to cancel the pole at z 5 0.995, leaving
the loop gain
LðzÞ 5 1:2417 3 1025
z 1 0:9934
ðz 2 1Þðz 2 0:9851Þ
The system root locus of Figure 6.28 shows that the closed-loop poles are close to the
unit circle at low gains, and the system is unstable at higher gains. Clearly, an additional
zero is needed to meet the design specification, and a PID controller is required. To make
the controller realizable, a pole must be added at z 5 0. The simplest design is then to cancel the pole closest to but not on the unit circle, giving the loop gain
LðzÞ 5 1:2417 3 1025
z 1 0:9934
zðz 2 1Þ
Adding the zero to the transfer function, we obtain the root locus of Figure 6.29 and
select a gain of 20,200. The corresponding time response is shown in Figure 6.30. The
time response shows less than 5% overshoot with a fast time response that meets all
design specifications. The design is better than that of Example 6.11, where the digital
controller was obtained via analog design.
Although it may be possible to improve the analog design to obtain better results than
those of Example 6.11, this requires trial and error as well as considerable experience. By
contrast, the digital design is obtained here directly in the z-domain without the need for
trial and error. This demonstrates that direct design in the z-domain using CAD tools can
be easier than indirect design.
6.5 Frequency response design
1
14
Imaginary Axis
251
0.8
0.1
0.6
0.2
0.3 System: l
0.4
0.5
0.6
0.7
0.8
0.4
0.2
188
Gain: 2.02e + 004
Pole: 0.361 + 0.325i
Damping: 0.701
Overshoot (%): 4.55
Frequency (rad/sec): 206
126
62.8
0.9
0
0
0.2
0.4
0.6
Real Axis
0.8
1
FIGURE 6.29
Root locus for digital PID control design by pole-zero cancellation.
System: cl
Peak amplitude: 1.04
Overshoot (%): 4.36
At time: 0.025
1
Amplitude
0.8
0.6
0.4
0.2
0
0
0.01
0.02
0.03
Time s
0.04
0.05
0.06
FIGURE 6.30
Time step response for digital PID control.
6.5 Frequency response design
Frequency response design approaches, especially design based on Bode plots,
are very popular in the design of analog control systems. They exploit the fact
that the Bode plots of rational loop gains (i.e., ones with no delays) can be
approximated with straight lines. Further, if the transfer function is minimum
205
206
CHAPTER 6 Digital Control System Design
phase, the phase can be determined from the plot of the magnitude, and this
allows us to simplify the design of a compensator that provides specified stability
properties. These specifications are typically given in terms of the phase margin
and the gain margin. As with other design approaches, the design is greatly simplified by the availability of CAD packages.
Unfortunately, discrete transfer functions are not rational functions in jω
because the frequency is introduced through the substitution z 5 ejωT in the
z-transfer function (see Section 2.8). Hence, the simplification provided by methods
based on Bode plots for analog systems is lost. A solution to this problem is to
bilinearly transform the z-plane into a new plane, called the w-plane, where the
corresponding transfer function is rational and where the Bode approximation is
valid. For this purpose, we recall the bilinear transformation
w5c
z21
z11
where
c5
2
T
(6.36)
from Section 6.3.3, which maps points in the LHP into points inside the unit circle. To transform the inside of the unit circle to the LHP, we use the inverse bilinear transformation
z5
11
12
wT
2
wT
2
(6.37)
This transforms the transfer function of a system from the z-plane to the
w-plane. The w-plane is a complex plane whose imaginary part is denoted by v.
To express the relationship between the frequency ω in the s-plane and the frequency v in the w-plane, we let s 5 jω and therefore z 5 ejωT. Substituting in
(6.36) and using steps similar to those used to derive (6.25), we have
2
ωT
w 5 jv 5 j tan
T
2
(6.38)
From (6.38), as ω varies in the interval [0, π/T], z moves on the unit circle and
v goes from 0 to infinity. This implies that there is a distortion or warping of the
frequency scale between v and ω (see Section 6.3.3). This distortion is significant,
especially at high frequencies. However, if ω{ωs/2 5 π/T, we have from (6.38)
that ω v and the distortion is negligible.
In addition to the problem of frequency distortion, the transformed transfer
function G(w) has two characteristics that can complicate the design: (1) the
transfer function will always have a pole-zero deficit of zero (i.e., the same number of poles as zeros) and (2) the bilinear transformation (6.36) can introduce
RHP zeros and result in a nonminimum phase system.
Nonminimum phase systems limit the achievable performance. For example,
because some root locus branches start at the open-loop pole and end at zeros, the
presence of RHP zeros limits the stable range of gains K. Thus, attempting to
reduce the steady-state error or to speed up the response by increasing the system
6.5 Frequency response design
gain can lead to instability. These limitations clearly make w-plane design challenging. Examples 6.16, 6.17, and 6.18 illustrate w-plane design and how to overcome its limitations.
We summarize the steps for controller design in the w-plane in the following
procedure.
PROCEDURE 6.2
1. Select a sampling period and obtain the transfer function GZAS(z) of the discretized
process.
2. Transform GZAS(z) into G(w) using (6.37).
3. Draw the Bode plot of G(jv), and use analog frequency response methods to design a
controller C(w) that satisfies the frequency domain specifications.
4. Transform the controller back into the z-plane by means of (6.36), thus determining
C(z).
5. Verify that the performance obtained is satisfactory.
In controller design problems, the specifications are often given in terms of
the step response of the closed-loop system, such as the settling time and the percentage overshoot. As shown in Chapter 5, the percentage overshoot specification
yields the damping ratio ζ, which is then used with the settling time to obtain the
undamped natural frequency ωn. Thus, we need to obtain frequency response specifications from ζ and ωn to follow Procedure 6.2.
The relationship between the time domain criteria and the frequency domain
criteria is in general quite complicated. However, the relationship is much simpler
if the closed-loop system can be approximated by the second-order underdamped
transfer function
TðsÞ 5
ω2n
s2 1 2ςωn s 1 ω2n
(6.39)
The corresponding loop gain with unity feedback is given by
LðsÞ 5
ω2n
s2 1 2ςωn s
(6.40)
We substitute s 5 jω to obtain the corresponding frequency response and
equate the square of its magnitude to unity:
LðjωÞ2 5
1
51
ðω=ωn Þ4 1 4ζ 2 ðω=ωn Þ2
(6.41)
The magnitude of the loop gain, as well as its square, is unity at the gain crossover frequency. For the second-order underdamped case, we now have the relation
qffiffiffiffiffiffiffiffiffiffiffiffiffiffi
(6.42)
ωgc 5 ωn ½ 4ζ 4 11 22ζ 2 1=2
207
208
CHAPTER 6 Digital Control System Design
Next, we consider the phase margin and derive
21
PM 5 180 1 +Gðjωgc Þ 5 tan
2ζ
pffiffiffiffiffiffiffiffiffiffiffiffiffiffi
4
½ 4ζ 11 22ζ 2 1=2
!
The last expression can be approximated by
PM 100ζ
(6.43)
Equations (6.42) and (6.43) provide the transformations we need to obtain frequency domain specifications from step response specifications. Together with
Procedure 6.2, the equations allow us to design digital control systems using the
w-plane approach. The following three examples illustrate w-plane design using
Procedure 6.2.
EXAMPLE 6.16
Consider the cruise control system of Example 3.2, where the analog process is
GðsÞ 5
1
s11
Transform the corresponding GZAS(z) to the w-plane by considering both T 5 0.1 and
T 5 0.01. Evaluate the role of the sampling period by analyzing the corresponding Bode plots.
Solution
When T 5 0.1 we have
GZAS ðzÞ 5
0:09516
z 2 0:9048
and by applying (6.37), we obtain
G1 ðwÞ 5
2 0:05w 1 1
w11
Note that the transfer function G1(w) can be obtained using the MATLAB command
.. Gw 5 d2cðGzas; 'tustin'Þ
When T 5 0.01 we have
GZAS ðzÞ 5
0:00995
z 2 0:99
and, again by applying (6.37), we obtain
G2 ðwÞ 5
20:005w 1 1
w11
The Bode plots of G(s), G1(w), and G2(w) are shown in Figure 6.31. For both sampling
periods, the pole in the w-plane is in the same position as the pole in the s-plane.
However, both and G2(w) have a zero, whereas G(s) does not. This results in a big difference between the frequency response of the analog system and that of the digital systems
at the high frequencies. However, the influence of the zero on the system dynamics is
clearly more significant when the sampling period is smaller. Note that for both sampling
periods, distortion in the low-frequency range is negligible. For both systems, the gain as
6.5 Frequency response design
Magnitude (dB)
0
G1(w)
–20
G(s)
–40
G2(w)
–60
–80
Phase (deg)
0
–45
–90
–135
–180
G_2(w)
10–2
10–1
100
101
102
Frequency (rad/sec)
G_1(w)
103
104
FIGURE 6.31
Bode plots for Example 6.16.
w goes to zero is unity, as is the DC gain of the analog system. This is true for the choice
c 5 2/T in (6.36). Other possible choices are not considered because they do not yield DC
gain equality.
EXAMPLE 6.17
Consider a DC motor speed control system where the (type 0) analog plant has the transfer
function
GðsÞ 5
1
ðs 1 1Þðs 1 10Þ
Design a digital controller by using frequency response methods to obtain zero steadystate error due to a unit step, an overshoot less than 10%, and a settling time of about 1 s.
Solution
From the given specification, we have that the controller in the w-plane must contain a
pole at the origin. For 10% overshoot, we calculate the damping ratio as
lnð0:1Þ
q
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ffi 0:6
ζ5 lnð0:1Þ2 1 π2
Using the approximate expression (6.43), the phase margin is about 100 times the
damping ratio of the closed-loop system, and the required phase margin is about 60 .
For a settling time of 1 second, we calculate the undamped natural frequency
ωn 5
4
6:7 rad=s
ζTs
209
CHAPTER 6 Digital Control System Design
Using (6.42), we obtain the gain crossover ωgc 5 4.8 rad/s.
A suitable sampling period for the selected dynamics is T 5 0.02 s (see Example 6.9).
The discretized process is then determined as
GðsÞ
z 1 0:9293
5 1:8604 3 1024
GZAS ðzÞ 5 ð1 2 z21 ÞZ
s
ðz 2 0:8187Þðz 2 0:9802Þ
Using (6.37), we obtain the w-plane transfer function
GðwÞ 5
23:6519U1026 ðw 1 2729Þðw 2 100Þ
ðw 1 9:967Þðw 1 1Þ
Note the two additional zeros (with respect to G(s)) do not significantly influence the
system dynamics in the range of frequencies of interest for the design. The two poles
are virtually in the same position as the poles of G(s). The simplest design that meets the
desired specifications is to insert a pole at the origin, to cancel the dominant pole at 2 1,
and to increase the gain until the required gain crossover frequency is attained. Thus, the
resulting controller transfer function is
CðwÞ 5 54
w11
w
The Bode diagram of the loop transfer function C(w)G(w), together with the Bode diagram of G(w), is shown in Figure 6.32. The figure also shows the phase and gain margins.
By transforming the controller back to the z-plane using (6.36), we obtain
CðzÞ 5
54:54z 2 53:46
z21
GM = 25.7 dB (at 32.2 rad/sec),
PM = 61.3 deg (at 4.86 rad/sec)
Magnitude (dB)
100
50
C(w)G(w)
0
–50
G(w)
–100
–150
Phase (deg)
210
0
–45
–90
–135
–180
–225
–270
10–2 10–1
100
102
103
101
Frequency (rad/sec)
FIGURE 6.32
Bode plots of C(w)G(w) and G(w) of Example 6.17.
104
105
6.5 Frequency response design
System: Gcl
Peak amplitude: 1.07
Overshoot (%): 7.17
At time (sec): 0.56
1.2
System: Gcl
Settling Time (sec): 0.818
Amplitude
1
0.8
0.6
0.4
0.2
0
0
0.2
0.4
0.6
Time s
0.8
1
1.2
FIGURE 6.33
Closed-loop step response for Example 6.17.
The transfer function C(z) can be obtained using the MATLAB command
.. Cz 5 c2dðCw; 0:02; 'matched'Þ
The corresponding discretized closed-loop step response is plotted in Figure 6.33 and
clearly meets the design specifications.
EXAMPLE 6.18
Consider the DC motor speed control system with transfer function
GðsÞ 5
1
ðs 1 1Þðs 1 3Þ
Design a digital controller using frequency response methods to obtain (zero steadystate error due to a unit step, an overshoot less than 10%, and a settling time of about
1 s. Use a sampling period T 5 0.2 s.
Solution
As in Example 6.17, (1) the given steady-state error specification requires a controller
with a pole at the origin, (2) the percentage overshoot specification requires a phase margin of about 60 , and (3) the settling time yields a gain crossover frequency of about
5 rad/s. The transfer function of the system with DAC and ADC is
GðsÞ
GZAS ðzÞ 5 ð1 2 z21 Þ Z
s
z 1 0:7661
5 0:015437
ðz 2 0:8187Þðz 2 0:5488Þ
211
CHAPTER 6 Digital Control System Design
GM = 9.7 dB (at 18.2 rad/sec),
PM = 59.9 deg (at 3.99 rad/sec)
Magnitude (dB)
40
20
C(w)G(w)
0
G(w)
–20
–40
–60
0
Phase (deg)
212
–45
–90
–135
–180
–225
10–1
100
101
102
Frequency (rad/sec)
103
104
FIGURE 6.34
Bode plots of C(w)G(w) and G(w) for Example 6.18.
Transforming to the w-plane using (6.37), we obtain
GðwÞ 5
212:819U1024 ðw 1 75:5Þðw 2 10Þ
ðw 1 2:913Þðw 1 0:9967Þ
Note again that the two poles are virtually in the same locations as the poles of G(s).
Here the RHP zero at w 5 10 must be taken into account in the design because the
required gain crossover frequency is about 5 rad/s. To achieve the required phase margin,
both poles must be canceled with two controller zeros. For a realizable controller, we need
at least two controller poles. In addition to the pole at the origin, we select a high-frequency controller pole so as not to impact the frequency response in the vicinity of the
crossover frequency. Next, we adjust the system gain to achieve the desired specifications.
As a first attempt, we select a high-frequency pole in w 5 220 and increase the gain to
78 for a gain crossover frequency of 4 rad/s with a phase margin of 60 degrees. The controller is therefore
CðwÞ 5 78
ðw 1 2:913Þðw 1 0:9967Þ
wðw 1 20Þ
The corresponding open-loop Bode plot is shown in Figure 6.34 together with the Bode
plot of G(w). Transforming the controller back to the z-plane using (6.36), we obtain
CðzÞ 5 CðwÞw52 ½z21 5
T z11
36:9201z2 2 50:4902z 1 16:5897
ðz 2 1Þðz 2 0:3333Þ
The resulting discretized closed-loop step response, which satisfies the given requirements, is plotted in Figure 6.35.
6.6 Direct control design
System: Gcl
Peak amplitude: 1.08
Overshoot (%): 8.09
At time (sec): 0.4
1.2
System: Gcl
Settling Time (sec): 0.718
Amplitude
1
0.8
0.6
0.4
0.2
0
0
0.2
0.4
0.6
0.8
Time s
1
1.2
1.4
FIGURE 6.35
Closed-loop step response for Example 6.18.
6.6 Direct control design
In some control applications, the desired transfer function of the closed-loop system is known from the design specification. For a particular system configuration,
it is possible to calculate the controller transfer function for a given plant from
the desired closed-loop transfer function. This approach to design is known as
synthesis. Clearly, the resulting controller must be realizable for this approach to
yield a useful design.
We consider the Figure 6.12 block diagram, with known closed-loop transfer
function Gcl(z). The controller transfer function C(z) can be computed analytically, starting from the expression of the desired closed-loop transfer function
Gcl(z):
Gcl ðzÞ 5
CðzÞGZAS ðzÞ
1 1 CðzÞGZAS ðzÞ
We solve for the controller transfer function
CðzÞ 5
1
Gcl ðzÞ
GZAS ðzÞ 1 2 Gcl ðzÞ
(6.44)
For the control system to be implementable, the controller must be causal and
must ensure the asymptotic stability of the closed-loop control system. For a
causal controller, the closed-loop transfer function Gcl(z) must have the same
pole-zero deficit as GZAS(z). In other words, the delay in Gcl(z) must be at least as
long as the delay in GZAS(z). We see this by examining the degree of the numerator and the degree of the denominator in (6.44).
213
214
CHAPTER 6 Digital Control System Design
Recall from Chapter 4 that if unstable pole-zero cancellation occurs, the system is input-output stable but not asymptotically stable. This is because the
response due to the initial conditions is unaffected by the zeros and is affected by
the unstable poles, even if they cancel with a zero. Hence, one must be careful
when designing closed-loop control systems to avoid unstable pole-zero cancellations. This implies that the set of zeros of Gcl(z) must include all the zeros of
GZAS(z) that are outside the unit circle. Suppose that the process has un
unstable pole z 5 z; jzj . 1—namely
GZAS ðzÞ 5
G1 ðzÞ
z2z
In view of equation (6.44), we avoid unstable pole-zero cancellation by
requiring
1 2 Gcl ðzÞ 5
1
1 1 CðzÞ Gz 21 ðzÞz
5
z2z
z 2 z 1 CðzÞG1 ðzÞ
In other words, z 5 z must be a zero of 1 2 Gcl(z).
An additional condition can be imposed to address steady-state accuracy
requirements. In particular, if zero steady-state error due to a step input is required,
the condition must be
Gcl ð1Þ 5 1
This condition is easily obtained by applying the Final Value theorem (see
Section 2.8.1). The proof is left as an exercise for the reader.
Summarizing, the conditions required for the choice of Gcl(z) are as follows:
•
•
•
•
Gcl(z) must have at least the same pole-zero deficit as GZAS(z) (causality).
Gcl(z) must contain as zeros all the zeros of GZAS(z) that are outside the unit
circle (stability).
The zeros of 1 2 Gcl(z) must include all the poles of GZAS(z) that are outside
the unit circle (stability).
Gcl(1) 5 1 (zero steady-state error).
The choice of a suitable closed-loop transfer function is clearly the main
obstacle in the application of the direct design method. The correct choice of
closed-loop poles and zeros to meet the design requirements can be difficult if
there are additional constraints on the control variable because of actuator limitations. Further, the performance of the control system relies heavily on an accurate
process model.
To partially address the first problem, instead of directly selecting the poles
and the zeros of the (first- or second-order) closed-loop system directly in the
z-plane, a continuous-time closed-loop system can be specified based on desired
properties of the closed-loop system, and then Gcl(z) is obtained by applying
6.6 Direct control design
pole-zero matching. The following ad hoc procedure often yields the desired
transfer function.
PROCEDURE 6.3
1. Select the desired settling time Ts (and the desired maximum overshoot).
2. Select a suitable continuous-time closed-loop first-order or second-order closed-loop
system with unit gain.
3. Obtain Gcl(z) by converting the s-plane pole location to the z-plane pole location
using pole-zero matching.
4. Verify that Gcl(z) meets the conditions for causality, stability, and steady-state error.
If not, modify Gcl(z) until the conditions are met.
EXAMPLE 6.19
Design a digital controller for a DC motor speed control system where the (type 0) analog
plant has the transfer function
GðsÞ 5
1
ðs 1 1Þðs 1 10Þ
to obtain zero steady-state error due to a unit step and a settling time of about 4 s. The
sampling period is chosen as T 5 0.02 s.
Solution
As in Example 6.9, the discretized process transfer function is
GðsÞ
z 1 0:9293
5 1:8604 3 1024
GZAS ðzÞ 5 ð1 2 z21 ÞZ
s
ðz 2 0:8187Þðz 2 0:9802Þ
Note that there are no poles or zeros outside the unit circle and the pole-zero deficit is
unity.
We choose the desired continuous-time closed-loop transfer function as second-order
with damping ratio 0.88, undamped natural frequency 1.15 rad/s, and with unity gain
to meet the settling time and steady-state error specifications. The resulting transfer
function is
Gcl ðsÞ 5
1:322
s2 1 2:024s 1 1:322
The desired closed-loop transfer function is obtained using pole-zero matching:
Gcl ðzÞ 5 0:25921U1023
z11
z2 2 1:96z 1 0:9603
By applying (6.44), we have
CðzÞ 5
1:3932ðz 2 0:8187Þðz 2 0:9802Þðz 1 1Þ
ðz 2 1Þðz 1 0:9293Þðz 2 0:9601Þ
We observe that stable pole-zero cancellation has occurred and that an integral term is
present in the controller, as expected, because the zero steady-state error condition has
been addressed. The closed-loop step response is shown in Figure 6.36.
215
CHAPTER 6 Digital Control System Design
1
System: gcl
Settling Time (sec): 3.71
0.8
Amplitude
216
0.6
0.4
0.2
0
0
1
2
3
Time s
4
5
6
FIGURE 6.36
Step response for the direct design of Example 6.19.
EXAMPLE 6.20
Design a digital controller for the type 0 analog plant
GðsÞ 5
20:16738ðs 2 9:307Þðs 1 6:933Þ
ðs2 1 0:3311s 1 9Þ
to obtain zero steady-state error due to a unit step, a damping ratio of 0.7, and a settling
time of about 1 s. The sampling period is chosen as T 5 0.01 s.
Solution
The discretized process transfer function is
GðsÞ
2 0:16738ðz 2 1:095Þðz 2 0:9319Þ
5
GZAS ðzÞ 5 ð1 2 z21 ÞZ
s
z2 2 1:996z 1 0:9967
Note that there is a zero outside the unit circle and this must be included in the
desired closed-loop transfer function, which is therefore selected as
Gcl ðzÞ 5
Kðz 2 1:095Þ
z2 2 1:81z 1 0:8269
The closed-loop poles have been selected as in Example 6.9, and the value
K 5 20.17789 results from solving the equation Gcl(1) 5 1. By applying (6.44), we have
CðzÞ 5
1:0628ðz2 2 1:81z 1 0:8269Þ
ðz 2 1Þðz 2 0:9319Þðz 2 0:6321Þ
The stable pole-zero cancellation is evident again, as well as the presence of the pole at
z 5 1. The digital closed-loop step response is shown in Figure 6.37. The undershoot is due
to the unstable zero that cannot be modified by the direct design method. The settling time
is less than the required value, as the presence of the zero has not been considered when
selecting the closed-loop poles based on the required damping ratio and settling time.
6.6 Direct control design
1.2
1
System: gcl
Settling Time (sec): 0.489
0.8
Amplitude
0.6
0.4
0.2
0
–0.2
–0.4
–0.6
0
0.1
0.2
0.3
0.4
Time s
0.5
0.6
0.7
FIGURE 6.37
Step response for the direct design of Example 6.20.
EXAMPLE 6.21
Design a digital controller for the type 0 analog plant
GðsÞ 5
1
e25s
10s 1 1
to obtain zero steady-state error that results from a unit step and a settling time of about
10 s with no overshoot. The sampling period is chosen as T 5 1 s.
Solution
The discretized process transfer function is
GZAS ðzÞ 5 ð1 2 z21 Þ Z
GðsÞ
0:09516 25
5
z
s
z 2 0:9048
To meet the causality requirements, a delay of five sampling periods must be included
in the desired closed-loop transfer function. Then the settling time of 10 s (including the
time delay) is achieved by considering a closed-loop transfer function with a pole in
z 5 0.5—namely, by selecting
Gcl ðzÞ 5
K
z25
z 2 0:5
Setting Gcl(1) 5 1 yields K 5 0.5, and then applying (6.44), we have
CðzÞ 5
5:2543ðz 2 0:9048Þz5
z6 2 0:5z5 2 0:5
The resulting closed-loop step response is shown in Figure 6.38.
217
CHAPTER 6 Digital Control System Design
1
0.8
Amplitude
218
0.6
0.4
0.2
0
0
2
4
6
8
10 12
Time s
14
16
18
20
FIGURE 6.38
Step response for the direct design of Example 6.21.
6.7 Finite settling time design
Continuous-time systems can only reach the desired output asymptotically after
an infinite time period. By contrast, digital control systems can be designed to settle at the reference output after a finite time period and follow it exactly thereafter. By following the direct control design method described in Section 6.6, if all
the poles and zeros of the discrete-time process are inside the unit circle, an
attractive choice is to select
Gcl ðzÞ 5 z2k
where k must be greater than or equal to the intrinsic delay of the discretized process—namely, the difference between the degree of the denominator and the
degree of the numerator of the discrete process transfer function.
Disregarding the time delay, the definition implies that a unit step is tracked
perfectly starting at the first sampling point. From (6.44) we have the deadbeat
controller
1
z2k
(6.45)
CðzÞ 5
GZAS ðzÞ 1 2 z2k
In this case, the only design parameter is the sampling period T, and the overall control system design is very simple. However, finite settling time designs
may exhibit undesirable intersample behavior (oscillations) because the control is
unchanged between two consecutive sampling points. Further, the control variable
can easily assume values that may cause saturation of the DAC or exceed the limits of the actuator in a physical system, resulting in unacceptable system behavior.
6.7 Finite settling time design
The behavior of finite settling time designs such as the deadbeat controller must
therefore be carefully checked before implementation.
EXAMPLE 6.22
Design a deadbeat controller for a DC motor speed control system where the (type 0)
analog plant has transfer the function
GðsÞ 5
1
ðs 1 1Þðs 1 10Þ
and the sampling period is initially chosen as T 5 0.02 s. Redesign the controller with T 5 0.1 s.
Solution
Because the discretized process transfer function is
GðsÞ
GZAS ðzÞ 5 ð1 2 z21 Þ Z
s
5 1:8604 3 1024
z 1 0:9293
ðz 2 0:8187Þðz 2 0:9802Þ
we have no poles and zeros outside or on the unit circle and a deadbeat controller can
therefore be designed by setting
Gcl ðzÞ 5 z21
Note that the difference between the order of the denominator and the order of the
numerator is one. By applying (6.45), we have
CðzÞ 5
5375:0533ðz 2 0:9802Þðz 2 0:8187Þ
ðz 2 1Þðz 1 0:9293Þ
The resulting sampled and analog closed-loop step response is shown in Figure 6.39,
whereas the corresponding control variable is shown in Figure 6.40. It appears that, as
expected, the sampled process output attains its steady-state value after just one sample—namely, at time t 5 T 5 0.02 s, but between samples the output oscillates wildly and
the control variable assumes very high values. In other words, the oscillatory behavior of
the control variable causes unacceptable intersample oscillations.
This can be ascertained analytically by considering the block diagram shown in
Figure 6.41. Recall that the transfer function of the zero-order hold is
GZOH ðsÞ 5
1 2 e2sT
s
Using simple block diagram manipulation, we obtain
E ðsÞ 5
R ðsÞ
1 1 ð1 2 e2sT ÞC ðsÞ GðsÞ
s
R ðsÞ
5
1 1 C ðsÞGZAS ðsÞ
(6.46)
Thus, the output is
YðsÞ 5
1 2 e2sT
R ðsÞ
GðsÞC ðsÞ
s
1 1 C ðsÞGZAS ðsÞ
(6.47)
219
CHAPTER 6 Digital Control System Design
Process Output
1.5
1
0.5
0
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
Time s
FIGURE 6.39
Sampled and analog step response for the deadbeat control of Example 6.22 (T 5 0.02 s).
× 104
1
0.8
0.6
0.4
Control Variable
220
0.2
0
–0.2
–0.4
–0.6
–0.8
–1
0
0.2
0.4
0.6
0.8
1
1.2
Time s
1.4
1.6
1.8
FIGURE 6.40
Control variable for the deadbeat control of Example 6.22 (T 5 0.02 s).
2
6.7 Finite settling time design
R(s) T
R*(s)
+
U*(z)
E*(s)
Y(s)
GZOH(s)
C*(s)
G(s)
T
−
Y*(s)
T
FIGURE 6.41
Block diagram for finite settling time design with analog output.
Equations (6.46) and (6.47) are quite general and apply to any system described by
the block diagram shown in Figure 6.41. To obtain the specific output pertaining to our
example, we substitute z 5 esT in the relevant expressions to obtain
GZAS ðsÞ 5 1:8604 3 1024 e2sT
C ðsÞ 5 5375:0533
R ðsÞ 5
1 1 0:9293e2sT
ð1 2 0:8187e2sT Þð1 2 0:9802e2sT Þ
ð1 2 0:9802e2sT Þð1 2 0:8187e2sT Þ
ð1 2 e2sT Þð1 1 0:9293e2sT Þ
1
1 2 e2sT
Then substituting in (6.47) and simplifying gives
YðsÞ 5 5375:0533
ð1 2 0:9802e2sT Þð1 2 0:8187e2sT Þ
sðs 1 1Þðs 1 10Þð1 1 0:9293e2sT Þ
Expanding the denominator using the identity
ð110:9293e2sT Þ21 5 1 1ð2 0:9293e2sT Þ 1 ð20:9293e2sT Þ2 1 ð20:9293e2sT Þ3 1 . . .
5 1 2 0:9293e2sT 1 0:8636e22sT 2 0:8025e23sT 1 . . .
we obtain
YðsÞ 5
5375:0533
ð1 2 2:782e2sT 1 3:3378e22sT
sðs 1 1Þðs 1 10Þ
2 2:2993e23sT 1 0:6930e24sT 1 . . .Þ
Finally, we inverse Laplace transform to obtain the analog output
0
1
1
1 2t
1 210t A
@
2 e 1 e
yðtÞ 5 5375:0533
1ðtÞ 1
10 9
90
5375:0533ð2 0:2728 1 0:3031e2ðt2TÞ 2 0:0303e210ðt2TÞ Þ1ðt 2 TÞ 1
5375:0533ð0:3338 2 0:3709e2ðt22TÞ 1 0:0371e210ðt22TÞ Þ1ðt 2 2TÞ 1
5375:0533ð2 0:2299 1 0:2555e2ðt23TÞ 2 0:0255e210ðt23TÞ Þ1ðt 2 3TÞ 1
5375:0533ð0:0693 2 0:0770e2ðt24TÞ 1 0:0077e210ðt24TÞ Þ1ðt 2 4TÞ 1 . . .
where 1(t) is the unit step function. It is easy to evaluate that, at the sampling points, we
have y(0.02) 5 y(0.04) 5 y(0.06) 5 y(0.08) 5 . . . 5 1, but between samples the output
221
CHAPTER 6 Digital Control System Design
1.4
1.2
1
Process Output
222
0.8
0.6
0.4
0.2
0
0
0.2
0.4
0.6
0.8
1
1.2
Time s
1.4
1.6
1.8
2
FIGURE 6.42
Sampled and analog step response for the deadbeat control of Example 6.22 (T 5 0.1 s).
oscillates wildly, as shown in Figure 6.39. To reduce intersample oscillations, we set
T 5 0.1 s and obtain the transfer function
GðsÞ
z 1 0:6945
5 35:501 3 1024
GZAS ðzÞ 5 ð1 2 z21 Þ Z
s
ðz 2 0:9048Þðz 2 0:3679Þ
For Gcl(z) 5 z21, we obtain
CðzÞ 5
281:6855ðz 2 0:9048Þðz 2 0:3679Þ
ðz 2 1Þðz 1 0:6945Þ
The resulting sampled and analog closed-loop step responses are shown in Figure 6.42,
whereas the corresponding control variable is shown in Figure 6.43. The importance of the
sampling period selection in the deadbeat controller design is obvious. Note that the oscillations and the amplitude of the control signal are both reduced, but it is also evident that the
intersample oscillations cannot be avoided completely with this approach.
To avoid intersample oscillations, we maintain the control variable constant
after n samples, where n is the degree of the denominator of the discretized process. This is done at the expense of achieving the minimum settling time (by considering the sampled process output) as in the previous examples. Considering the
control scheme shown in Figure 6.12, we have
UðzÞ 5
YðzÞ
YðzÞ RðzÞ
RðzÞ
5
5 Gcl ðzÞ
GZAS ðzÞ
RðzÞ GZAS ðzÞ
GZAS ðzÞ
(6.48)
If we solve this equation with the constraints Gcl(1) 5 1, we obtain the expression of U(z), or, alternatively, we obtain the expression of Gcl(z), which yields the
6.7 Finite settling time design
300
Control Variable
200
100
0
–100
–200
–300
0
0.2
0.4
0.6
0.8
1
1.2
Time s
1.4
1.6
1.8
2
FIGURE 6.43
Control variable for the deadbeat control of Example 6.22 (T 5 0.1 s).
expression of the controller C(z) by applying (6.44). Obviously, the overall design
requires the a priori specification of the reference signal to be tracked. By considering a step signal for the sake of simplicity, a ripple-free (that is, without intersample oscillations) response is achieved if, for any k $ n, we have:
• e(kT) 5 0
• u(kT) 5 cost
By denoting the loop transfer function as L(z) 5 C(z)GZAS(z) 5 NL(z)/DL(z), it
can be demonstrated that these two conditions are satisfied if and only if
NL ðzÞ 1 DL ðzÞ 5 zk
(6.49)
• DL(z) has a root at z 5 1
• There are no pole-zero cancellations between C(z) and GZAS(z)
If we consider a process without poles and zeros outside the unit circle, if the
system has not a pole at z 5 1, the controller can be determined by considering
GZAS ðzÞ 5
an21 zn21 1 ? 1 a0
DðzÞ
and by expressing the controller as
CðzÞ 5 K
ðz 2 1Þðzn21
DðzÞ
1 bn22 zn22 ? 1 b0 Þ
(6.50)
223
224
CHAPTER 6 Digital Control System Design
where the coefficients K, bn-1, . . ., b0 (and an-1, . . ., a0) are determined by solving
(6.49)—that is, by solving
an21 zn21 1 ? 1 a0 1ðz 2 1Þðzn21 1 ? 1 b0 Þ 5 zk
If the process has a pole at z 5 1, the same reasoning is applied with
an21 zn21 1 ? 1 a0
ðz 2 1ÞDðzÞ
GZAS ðzÞ 5
and
CðzÞ 5 K
DðzÞ
zn21 1 bn22 zn22 ? 1 b0
(6.51)
EXAMPLE 6.23
Design a ripple-free deadbeat controller for the type 1 vehicle positioning system of
Example 3.3 with transfer function
GðsÞ 5
1
sðs 1 1Þ
The sampling period is chosen as T 5 0.1.
Solution
The discretized process transfer function is
GðsÞ
s
0:0048374ð1 1 0:9672z21 Þz21
5
ð1 2 z21 Þð1 2 0:9048z21 Þ
GZAS ðzÞ 5 ð1 2 z21 ÞZ
Using the z-transform of the step reference signal
RðzÞ 5
1
1 2 z21
we have
UðzÞ 5 Gcl ðzÞ
RðzÞ
GZAS ðzÞ
5 Gcl ðzÞU
206:7218ð1 2 0:9048z21 Þ
z21 ð1 1 0:9672z21 Þ
(6.52)
Because the process is of type 1, we require that the control variable be zero after two
samples (note that n 5 2)—namely, that
UðzÞ 5 a0 1 a1 z21
(6.53)
A solution of equation (6.52) can be obtained if Gcl(z) cancels the denominator of the
second term in its right hand side. Thus, if we select the closed-loop transfer function
Gcl ðzÞ 5 KUz21 ð1 1 0:9672z21 Þ
6.7 Finite settling time design
1
Process Output
0.8
0.6
0.4
0.2
0
0
0.1
0.2
0.3
0.4
0.5 0.6
Time s
0.7
0.8
0.9
1
FIGURE 6.44
Sampled and analog step response for the deadbeat control of Example 6.23.
we have the control
UðzÞ 5 KU206:7218ð1 2 0:9048z21 Þ
The gain value K 5 0.5083 is found by requiring that Gcl(1) 5 1 for zero steady-state
error due to step. From (6.44), we have the controller
CðzÞ 5
105:1z 2 95:08
z 1 0:4917
The resulting sampled and analog closed-loop step response is shown in Figure 6.44,
whereas the corresponding control variable is shown in Figure 6.45. Note that the control
variable is constant after the second sample and that there is no intersample ripple.
EXAMPLE 6.24
Design a ripple-free deadbeat controller for a DC motor speed control system where the
(type 0) analog plant has the transfer function
GðsÞ 5
1
ðs 1 1Þðs 1 10Þ
The sampling period is chosen as T 5 0.1 s.
Solution
For a sampling period of 0.01, the discretized process transfer function is
GðsÞ
GZAS ðzÞ 5 ð1 2 z21 Þ Z
s
0:00355z 1 0:002465
5 2
z 2 1:273z 1 0:3329
225
CHAPTER 6 Digital Control System Design
150
100
Control Variable
226
50
0
–50
–100
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Time s
FIGURE 6.45
Control variable for the deadbeat control of Example 6.23.
As the process has no pole at the origin, the controller transfer function can be written
as (see (6.50))
CðzÞ 5 K
z2 2 1:273z 1 0:3329
ðz 2 1Þðz 1 b0 Þ
and solving (6.49) implies the solution of the system
0:00355K 2 1 1 b0 5 0
0:002465K 2 b0 5 0
which yields b0 5 0.4099 and K 5 166.2352—that is,
CðzÞ 5
166:2352ðz 2 0:9048Þðz 2 0:3679Þ
ðz 2 1Þðz 1 0:4099Þ
The resulting sampled and analog closed-loop step responses are shown in
Figure 6.46, whereas the corresponding control variable is shown in Figure 6.47. As in
Example 6.23, there is no intersample ripple.
If the process has poles or zeros on or outside the unit circle, the finite settling time
design procedure must be modified to include the constraints on the choice of Gcl(z) outlined in Section 6.6. Let the plant transfer function have the form
h
L ð1 2 zi z21 Þ
GZAS ðzÞ 5
i51
l
G1 ðzÞ
L ð1 2 pj z21 Þ
j51
where zi, i 5 1, . . ., h and pj, j 5 1, . . ., h are the zeros and the poles of Gcl(z) that are on or
outside the unit circle.
6.7 Finite settling time design
1
Process Output
0.8
0.6
0.4
0.2
0
0
0.1
0.2
0.3
0.4
0.5 0.6
Time s
0.7
0.8
0.9
1
FIGURE 6.46
Sampled and analog step response for the deadbeat control of Example 6.24.
200
Control Variable
150
100
50
0
–50
0
0.1
0.2
0.3
0.4
0.5 0.6
Time s
FIGURE 6.47
Control variable for the deadbeat control of Example 6.24.
0.7
0.8
0.9
1
227
228
CHAPTER 6 Digital Control System Design
For the closed-loop transfer function Gcl(z) to include all the zeros of GZAS(z) that are
outside the unit circle, it must be in the form
h
Gcl ðzÞ 5 L ð1 2 zi z21 Þz2nz ða0 1 a1 z21 1 ? 1 ad z2d Þ
i51
where nz is the zero-pole deficit of the process transfer function GZAS(z).
For 1 2 Gcl(z) to include all the poles of GZAS(z) that are on or outside the unit circle as
zeros, it must be in the form
l
1 2 Gcl ðzÞ 5 L ð1 2 pj z21 Þð12z21 Þno ð1 1 b1 z21 1 ? 1 bf z2f Þ
j51
where d, f, g, and no are parameters to be determined from constraints on the controller.
For zero steady-state error due to a step input, the control loop must include an integrator.
Therefore, we need to set no 5 1 if GZAS(z) has no poles at z 5 1 and no 5 0 otherwise.
Parameters d and f can be found, together with the degree g of Gcl(z), in order to meet
the following conditions:
•
•
The z21 polynomials Gcl(z) and 1 2 Gcl(z) must be of the same order.
The total number of unknown coefficients ai and bi must be equal to the order of
Gcl(z)—that is, to the number of equations available by equating coefficients of z21.
From the first condition, we have
d 1 nz 1 h 5 g
and
f 1 no 1 l 5 g
while from the second condition, we have
f 1d115g
The constraints can be written in matrix form as
3
2
32 3 2
21
0
1
d
nz 1 h
4 0
21
1 54 f 5 5 4 no 1 l 5
21
1
1
21
g
(6.51)
We therefore have a system of three equations that can be easily solved for the three
unknowns d, f, and g. The coefficients ai and bi are determined by equating the coefficients of the terms of the same order in the equation
h
l
i51
j51
1 2 L ð1 2 zi z21 Þz2k ða0 1 a1 z21 1 ? 1 ad z2d Þ 5 L ð1 2 pj z21 Þð12z21 ÞP ð1 1 b1 z21 1 ?bf z2f Þ
(6.52)
EXAMPLE 6.25
Design a deadbeat controller for an isothermal chemical reactor whose transfer function is1
GðsÞ 5
with a sampling period T 5 0.01.
21:1354ðs 2 2:818Þ
ðs 1 2:672Þðs 1 2:047Þ
6.7 Finite settling time design
Solution
For a sampling period of 0.01, the discretized process transfer function is
GZAS ðzÞ 5
20:010931ðz 2 1:029Þ
ðz 2 0:9797Þðz 2 0:9736Þ
The process has one zero and no poles outside the unit circle, and therefore we have
h 5 1 and l 5 0. Further, we have nz 5 1, and we select no 5 1 because the process transfer
function has no pole at z 5 1. By solving (6.51), we obtain d 5 0, f 5 1, and g 5 2. Equation
(6.52) is then written as
1 2ð1 2 1:029z21 Þz21 a0 5 ð1 2 z21 Þð1 1 b1 z21 Þ
and solved for a0 5 9.4958 and b1 5 35.9844. The resulting closed-loop transfer function
Gcl ðzÞ 5 2 34:9844ð1 2 1:029z21 Þz21
is substituted in (6.44) to obtain the controller transfer function
CðzÞ 5
3200:3413ðz 2 0:9797Þðz 2 0:9736Þ
ðz 2 1Þðz 1 35:98Þ
We observe that the controller has a pole outside the unit circle. In fact, there is no
guarantee of the stability of the controller by using the finite settling time design if the
process has poles or zeros outside the unit circle. Further, the transient response can
exhibit unacceptable behavior, as in the step response shown in Figures 6.48 and 6.49,
where we can observe a large undershoot and large control magnitudes.
1
Bequette, B.W., 2003. Process Control—Modeling, Design, and Simulation. Prentice Hall, Upper Saddle
River, NJ.
5
0
Process Output
–5
–10
–15
–20
–25
–30
–35
0
0.005 0.01 0.015 0.02 0.025 0.03 0.035 0.04 0.045 0.05
Time s
FIGURE 6.48
Sampled and analog step response for the deadbeat control of Example 6.25.
229
CHAPTER 6 Digital Control System Design
4000
3000
2000
Control Variable
230
1000
0
–1000
–2000
–3000
–4000
0
0.005 0.01 0.015 0.02 0.025 0.03 0.035 0.04 0.045 0.05
Time s
FIGURE 6.49
Control variable for the deadbeat control of Example 6.25.
Resources
Jacquot, R.G., 1981. Modern Digital Control Systems. Marcel Dekker, New York.
Kuo, B.C., 1991. Automatic Control Systems. Prentice Hall, Englewood Cliffs, NJ.
Kuo, B.C., 1992. Digital Control Systems. Saunders, Fort Worth, TX.
Ogata, K., 1987. Digital Control Engineering. Prentice Hall, Englewood Cliffs, NJ.
Oppenheim, A.V., Schafer, R.W., 1975. Digital Signal Processing. Prentice Hall,
Englewood Cliffs, NJ.
Ragazzini, J.R., Franklin, G.F., 1958. Sampled-Data Control Systems. McGraw-Hill, New York.
PROBLEMS
6.1
Sketch the z-domain root locus, and find the critical gain for the following
systems:
K
(a) GðzÞ 5
z 2 0:4
K
(b) GðzÞ 5
ðz 1 0:9Þðz 2 0:9Þ
Kz
(c) GðzÞ 5
ðz 2 0:2Þðz 2 1Þ
Kðz 1 0:9Þ
(d) GðzÞ 5
ðz 2 0:2Þðz 2 0:8Þ
Problems
6.2
Prove that expression (6.6) describes a constant ωn contour in the z-plane.
6.3
Design proportional controllers for the systems of Problem 6.1 to meet the
following specifications where possible. If the design specification cannot
be met, explain why and suggest a more appropriate controller.
(a) A damping ratio of 0.7
(b) A steady-state error of 10% due to a unit step
(c) A steady-state error of 10% due to a unit ramp
6.4
Hold equivalence is a digital filter design approach that approximates an
analog filter using
CðzÞ 5
z21
Ca ðsÞ Z L 21
z
s
(a) Obtain the hold-equivalent digital filter for the PD, PI, and PID controllers. Modify the results as necessary to obtain a realizable filter
with finite frequency response at the folding frequency.
(b) Why are the filters obtained using hold equivalence always stable?
6.5
Consider the system
Ga ðsÞ 5
YðsÞ
1
5
UðsÞ ðs11Þ2
Use pole-zero matching to determine a digital filter approximation with a
sampling period of T 5 0.1. Verify that not including a digital filter zero at
unity results in a time delay by comparing the unit step response for the
matched filter with and without a zero at 2 1.
6.6
Consider the PI controller
Ca ðsÞ 5 10
ðs 1 1Þ
s
Use pole-zero matching to determine a digital approximation with a sampling period of T 5 0.1.
6.7
Show that the bilinear transformation of the PID controller expression
(5.20) yields expression (6.33).
6.8
Show that the bilinear transformation of a PD controller with a high-frequency pole that makes the controller transfer function proper does not
yield a pole at z 5 1.
6.9
Design digital controllers to meet the desired specifications for the systems
described in Problems 5.4, 5.7, and 5.8 by bilinearly transforming the analog designs.
231
232
CHAPTER 6 Digital Control System Design
6.10
Design a digital filter by applying the bilinear transformation to the analog
(Butterworth) filter
Ca ðsÞ 5
s2
1
pffiffiffi
1 2s 1 1
with T 5 0.1 s. Then apply prewarping at the 3-dB frequency.
6.11
Design a digital PID controller (with T 5 0.1) for the plant
1
e25s
10s 1 1
GðsÞ 5
by applying the Ziegler-Nichols tuning rules presented in Table 5.1.
6.12
Design digital controllers to meet the desired specifications for the systems
described in Problems 5.4, 5.7, and 5.8 in the z-domain directly.
6.13
In Example 4.9, we examined the closed-loop stability of the furnace temperature digital control system with proportional control and a sampling
period of 0.01 s. We obtained the z-transfer function
GZAS ðzÞ 5 1025
4:95z 1 4:901
z2 2 1:97z 1 0:9704
Design a controller for the system to obtain zero steady-state error due to a
step input without significant deterioration in the transient response.
6.14
Consider the DC motor position control system described in Example 3.6,
where the (type 1) analog plant has the transfer function
GðsÞ 5
1
sðs 1 1Þðs 1 10Þ
and design a digital controller by using frequency response methods to
obtain a settling time of about 1 and an overshoot less than 5%.
6.15
Use direct control design for the system described in Problem 5.7 (with
T 5 0.1) to design a controller for the transfer function
GðsÞ 5
1
ðs 1 1Þðs 1 5Þ
to obtain zero steady-state error due to step, a settling time of less than 2 s,
and an undamped natural frequency of 5 rad/s. Obtain the discretized and
the analog output. Then apply the designed controller to the system
GðsÞ 5
1
ðs 1 1Þðs 1 5Þð0:1s 1 1Þ
Computer Exercises
and obtain the discretized and the analog output to verify the robustness of
the control system.
6.16
Design a deadbeat controller for the system of Problem 5.7 to obtain perfect tracking of a unit step in minimum finite time. Obtain the analog output for the system, and compare your design to that obtained in Problem
5.7. Then apply the controller to the process
GðsÞ 5
1
ðs 1 1Þðs 1 5Þð0:1s 1 1Þ
to verify the robustness of the control system.
6.17
Find a solution for Problem 6.16 that avoids intersample ripple.
6.18
Design a deadbeat controller for a steam-drum-level system whose transfer
function is2
GðsÞ 5
0:25ð2s 1 1Þ
sð2s 1 1Þ
by selecting a sampling period T 5 0.01.
COMPUTER EXERCISES
6.19
Write a MATLAB function to plot a constant damped natural frequency
contour in the z-plane.
6.20
Write a MATLAB function to plot a time-constant contour in the z-plane.
6.21
Write a computer program that estimates a first-order-plus-dead-time transfer function with the tangent method and determines the digital PID parameters according to the Ziegler-Nichols formula. Apply the program to the
system
GðsÞ 5
1
ðs11Þ8
and simulate the response of the digital control system (with T 5 0.1)
when a set point step change and a load disturbance step are applied.
Compare the results with those of Exercise 5.14.
2
Bequette, B.W., 2003. Process Control—Modeling, Design, and Simulation. Prentice Hall, Upper
Saddle River, NJ.
233
234
CHAPTER 6 Digital Control System Design
6.22
To examine the effect of the sampling period on the relative stability and
transient response of a digital control system, consider the system
GðsÞ 5
1
ðs 1 1Þðs 1 5Þ
(a) Obtain the transfer function of the system, the root locus, and the critical gain for T 5 0.01 s, 0.05 s, 0.1 s.
(b) Obtain the step response for each system at a gain of 2.
(c) Discuss the effect of the sampling period on the transient response
and relative stability of the system based on your results from (a)
and (b).
CHAPTER
StateSpace Representation
7
OBJECTIVES
After completing this chapter, the reader will be able to do the following:
1. Obtain a statespace model from the system transfer function or differential
equation.
2. Determine a linearized model of a nonlinear system.
3. Determine the solution of linear (continuous-time and discrete-time) statespace
equations.
4. Determine an input-output representation starting from a statespace
representation.
5. Determine an equivalent statespace representation of a system by changing the
basis vectors.
In this chapter, we discuss an alternative system representation in terms of the
system state variables, known as the statespace representation or realization.
We examine the properties, advantages, and disadvantages of this representation.
We also show how to obtain an input-output representation from a statespace
representation. Obtaining a statespace representation from an input-output representation is further discussed in Chapter 8.
The term realization arises from the fact that this representation provides the
basis for implementing digital or analog filters. In addition, statespace realizations
can be used to develop powerful controller design methodologies. Thus, statespace
analysis is an important tool in the arsenal of today’s control system designer.
7.1 State variables
Linear continuous-time single-input-single-output (SISO) systems are typically
described by the inputoutput differential equation
dn y
dn21 y . . .
dy
1
a
1
1 a1 1 a0 y
n21
n
n21
dt
dt
dt
5 cn
dn u
dn21 u
du
1 c0 u
1 cn21 n21 1 . . . 1 c1
n
dt
dt
dt
(7.1)
235
236
CHAPTER 7 StateSpace Representation
where y is the system output, u is the system input, and ai, i 5 0, 1, . . ., n 2 1, cj,
j 5 0, 1,. . . , n are constants. The description is valid for time-varying systems
if the coefficients ai and cj are explicit functions of time. For a multi-inputmulti-output (MIMO) system, the representation is in terms of l input-output
differential equations of the form (7.1), where l is the number of outputs. The
representation can also be used for nonlinear systems if (7.1) is allowed to include
nonlinear terms.
The solution of the differential equation (7.1) requires knowledge of the
system input u(t) for the period of interest as well as a set of constant initial
conditions
yðt0 Þ; dyðt0 Þ=dt; . . . ; d n21 yðt0 Þ=dtn21
where the notation signifies that the derivatives are evaluated at the initial time t0.
The set of initial conditions is minimal in the sense that incomplete knowledge
of this set would prevent the complete solution of (7.1). On the other hand,
additional initial conditions are not needed to obtain the solution. The initial
conditions provide a summary of the history of the system up to the initial time.
This leads to the following definition.
DEFINITION 7.1: SYSTEM STATE
The state of a system is the minimal set of numbers {xi(t0), i 5 1, 2, . . . , n} needed together
with the input u(t), with t in the interval [t0, tf), to uniquely determine the behavior of the
system in the interval [t0, tf]. The number n is known as the order of the system.
As t increases, the state of the system evolves and each of the numbers xi(t)
becomes a time variable. These variables are known as the state variables. In
vector notation, the set of state variables form the state vector
2
xðtÞ 5 x1
x2
. . . xn
T
x1
3
6x 7
6 27
56 7
4^5
(7.2)
xn
The preceding equation follows standard notation in system theory where a column
vector is bolded and a row vector is indicated by transposing a column.
Statespace is an n-dimensional vector space where {xi(t), i 5 1, 2, . . . , n}
represent the coordinate axes. So for a second-order system, the statespace is
two-dimensional and is known as the state plane. For the special case where
the state variables are proportional to the derivatives of the output, the state
plane is called the phase plane and the state variables are called the phase
7.1 State variables
variables. Curves in statespace are known as the state trajectories, and a plot
of state trajectories in the plane is the state portrait (or phase portrait for the
phase plane).
EXAMPLE 7.1
Consider the equation of motion of a point mass m driven by a force f
m€y 5 f
where y is the displacement of the point mass. The solution of the differential equation is
given by
yðtÞ 5
ðt ðt
1
_ 0 Þt 1
f ðτÞdτ
yðt0 Þ 1 yðt
m
t0 t0
Clearly, a complete solution can only be obtained, given the force, if the two initial
_ 0 Þg are known. Hence, these constants define the state of the system
conditions fyðt0 Þ; yðt
at time t0, and the system is second order. The state variables are
x1 ðtÞ 5 yðtÞ
_
x2 ðtÞ 5 yðtÞ
and the state vector is
xðtÞ 5 ½ x1
x 2 T 5
x1
x2
The state variables are phase variables in this case, because the second is the derivative
of the first.
The state variables are governed by the two first-order differential equations
x_1 5 x2
x_2 5 u=m
where u 5 f. The first of the two equations follows from the definitions of the two state
variables. The second is obtained from the equation of motion for the point mass. The two
differential equations together with the algebraic expression
y 5 x1
are equivalent to the second-order differential equation because solving the first-order differential equations and then substituting in the algebraic expression yields the output y.
For a force satisfying the state feedback law
u
5 29x1 2 3x2
m
we have a second-order underdamped system with the solution depending only on
the initial conditions. The solutions for different initial conditions can be obtained by
repeatedly using the MATLAB commands lsim or initial. Each of these solutions yields
position and velocity data for a phase trajectory, which is a plot of velocity versus
position. A set of these trajectories corresponding to different initial states gives
237
CHAPTER 7 StateSpace Representation
3
2
1
x2
238
0
–1
–2
–3
–1.5
–1
–0.5
0
0.5
1
1.5
x1
FIGURE 7.1
Phase portrait for a point mass.
the phase portrait of Figure 7.1. The time variable does not appear explicitly in the
phase portrait and is an implicit parameter. Arrows indicate the direction of increasing time.
Note that the choice of state variables is not unique. For example, one could
use the displacement y and the sum of displacement and velocity as state variables
(see Example 7.20). This choice has no physical meaning but nevertheless satisfies
the definition of state variables. The freedom of choice is a general characteristic of
state equations and is not restricted to this example. It allows us to represent a system so as to reveal its characteristics more clearly and is explored in later sections.
7.2 Statespace representation
In Example 7.1, two first-order equations governing the state variables were obtained
from the second-order input-output differential equation and the definitions of the
state variables. These equations are known as state equations. In general, there
are n state equations for an nth-order system. State equations can be obtained for
state variables of systems described by input-output differential equations, with the
form of the equations depending on the nature of the system. For example, the
equations are time varying for time-varying systems and nonlinear for nonlinear
systems. State equations for linear time-invariant systems can also be obtained
from their transfer functions.
The algebraic equation expressing the output in terms of the state variables is
called the output equation. For multi-output systems, a separate output equation
is needed to define each output. The state and output equations together provide
7.2 Statespace representation
a complete representation for the system described by the differential equation,
which is known as the statespace representation. For linear systems, it is often
more convenient to write the state equations as a single matrix equation referred
to as the state equation. Similarly, the output equations can be combined in a
single-output equation in matrix form. The matrix form of the statespace representation is demonstrated in Example 7.2.
EXAMPLE 7.2
The statespace equations for the system of Example 7.1 in matrix form are
2 3
0
x_1
0 1 x1
6 7
5
1 4 1 5u
x_2
0 0 x2
m
x
y 5 ½ 1 0 1
x2
The general form of the statespace equations for linear systems is
x_ ðtÞ 5 AxðtÞ 1 BuðtÞ
yðtÞ 5 CxðtÞ 1 DuðtÞ
(7.3)
where x(t) is an n 3 1 real vector, u(t) is an m 3 1 real vector, and y(t) is an l 3 1
real vector. The matrices in the equations are
A 5 n 3 n state matrix
B 5 n 3 m input or control matrix
C 5 l 3 n output matrix
D 5 l 3 m direct transmission matrix
The orders of the matrices are dictated by the dimensions of the vectors and the
rules of vector-matrix multiplication. For example, in the single-input (SI) case,
B is a column matrix, and in the single-output (SO) case, both C and D are row
matrices. For the SISO case, D is a scalar. The entries of the matrices are constant
for time-invariant systems and functions of time for time-varying systems.
EXAMPLE 7.3
The following are examples of statespace equations for linear systems.
1. A third-order 2-input2-output (MIMO) linear time-invariant system:
2 3 2
32 3 2
3
x_1
1:1 0:3 21:5
x1
0:1 0
u
4 x_2 5 5 4 0:1 3:5
5
4
5
4
x2 1 0 1:1 5 1
2:2
u2
x_3
x3
0:4 2:4 21:1
1:0 1:0
2 3
x
1 2 0 4 15
y1
1 2 u1
5
x2 1
0 0 1
y2
0 1 u2
x3
239
240
CHAPTER 7 StateSpace Representation
2. A second-order 2-outputsingle-input (SIMO) linear time-varying system:
x_1
sinðtÞ cosðtÞ x1
0
5
1
u
x_2
x2
1
e22t
1
y1
1 0 x1
5
y2
0 1 x2
Here, the direct transmission matrix D is zero, and the input and output matrices are constant. But the system is time varying because the state matrix has some entries that are
functions of time.
7.2.1 Statespace representation in MATLAB
MATLAB has a special statespace representation obtained with the command
ss. However, some state commands only operate on one or more of the matrices
(A, B, C, D). To enter a matrix
1
1
A5
25 24
use the command
.. A 5 ½0; 1; 5; 4
If B, C, and D are similarly entered, we obtain the statespace quadruple p with
the command
.. p 5 ssðA; B; C; DÞ
We can also specify names for the input (torque, say) and output (position) using
the set command
.. setðp; 'inputn'; 'torque'; 'outputn'; 'position'Þ
7.2.2 Linear versus nonlinear statespace equations
It is important to remember that the form (7.3) is only valid for linear state equations.
Nonlinear state equations involve nonlinear functions and cannot be written in
terms of the matrix quadruple (A, B, C, D).
EXAMPLE 7.4
Obtain a statespace representation for the s-degree-of-freedom (s-D.O.F.) robotic manipulator from the equation of motion
_ q_ 1 gðqÞ 5 τ
MðqÞ€q 1 Vðq; qÞ
7.2 Statespace representation
where
q 5 vector of generalized coordinates
M(q) 5 s 3 s positive definite inertia matrix
_ 5 s 3 s matrix of velocity-related terms
Vðq; qÞ
g(q) 5 s 3 1 vector of gravitational terms
τ 5 vector of generalized forces
The output of the manipulator is the position vector q.
Solution
The system is of order 2s, as 2s initial conditions are required to completely determine the
solution. The most natural choice of state variables is the vector
_
x 5 colfx1 ; x2 g 5 colfq; qg
where col{.} denotes a column vector. The associated state equations are
" _ # "
#
# "
x2
0
x1
5 1 u
x_ 2
2M 21 ðx1 ÞfVðx1 ; x2 Þx2 1 gðx1 Þg
M21 ðx1 Þ
with the generalized force now denoted by the symbol u
The output equation is
y 5 x1
This equation is linear and can be written in the standard form
#
"
x1
y 5 Is 0s 3 s x2
EXAMPLE 7.5
Write the statespace equations for the 2-D.O.F. anthropomorphic manipulator in
Figure 7.2. The equations of motion of the manipulator are as in Example 7.4 with the
definitions
"
MðθÞ 5
ðm1 1 m2 Þl21 1 m2 l22 1 2m2 l1 l2 cosðθ2 Þ
m2 l22 1 m2 l1 l2 cosðθ2 Þ
m2 l22 1 m2 l1 l2 cosðθ2 Þ
#
m2 l22
"
#
2m2 l1 l2 sinðθ2 Þθ_ 2 ð2θ_ 1 1 θ_ 2 Þ
_
_
Vðθ; θÞθ 5
2
m2 l1 l2 sinðθ2 Þθ_ 1
gðθÞ 5
ðm1 1 m2 Þgl1 sinðθ1 Þ 1 m2 gl2 sinðθ1 1 θ2 Þ
m2 gl2 sinðθ1 1 θ2 Þ
where mi, i 5 1, 2 are the masses of the two links; li, i 5 1, 2 are their lengths; and g is
_ are the vectors of angular positions and angular
the acceleration due to gravity. ðθ; θÞ
velocities, respectively.
241
242
CHAPTER 7 StateSpace Representation
θ2
Link 2
θ1
Link 1
FIGURE 7.2
A 2-D.O.F. anthropomorphic manipulator.
Solution
The state equations can be written using the results of Example 7.4 as
3
x2
6 7
6
97
"
#
6 8
> 2m2 l1 l2 sinðθ2 Þθ_ 2 ð2θ_ 1 1 θ_ 2 Þ
>7
7
6 >
>
x_ 1
>
>
71 0 u
6
1
>
>
56 <
2
=
7
L
x_ 2
m2 l1 l2 sinðθ2 Þθ_ 1
7
6 2L
6 >
> 7
5
4 >
>
m
ð
1
m
Þgl
sinðθ
Þ
1
m
gl
sinðθ
1
θ
Þ
1
2
1
1
2
2
1
2
>
>
>
>
:
;
m2 gl2 sinðθ1 1 θ2 Þ
2
where
L5
"
#
2ðm2 l22 1 m2 l1 l2 cosðθ2 ÞÞ
m2 l22
1
detðDÞ 2ðm2 l22 1 m2 l1 l2 cosðθ2 ÞÞ ðm1 1 m2 Þl21 1 m2 l22 1 2m2 l1 l2 cosðθ2 Þ
_
x 5 colfx1 ; x2 g 5 colfθ; θg
The general form of nonlinear statespace equations is
x_ 5 fðx; uÞ
y 5 gðx; uÞ
(7.4)
where f(.) (n 3 1) and g(.) (l 3 1) are vectors of functions satisfying mathematical
conditions that guarantee the existence and uniqueness of solution. But a form
that is often encountered in practice and includes the equations of robotic manipulators is
x_ 5 fðxÞ 1 BðxÞu
y 5 gðxÞ 1 DðxÞu
(7.5)
7.3 Linearization of nonlinear state equations
The state equation (7.5) is said to be affine in the control because the RHS is
affine (includes a constant vector) for constant x.
7.3 Linearization of nonlinear state equations
Nonlinear state equations of the form (7.4) or (7.5) can be approximated by linear
state equations of the form (7.3) for small ranges of the control and state variables.
The linear equations are based on the first-order approximation
f ðxÞ 5 f ðx0 Þ 1
df ∆x 1 Oð∆2 xÞ
dx x0
(7.6)
where x0 is a constant and ∆x 5 x 2 x0 is a perturbation from the constant.
The error associated with the approximation is of order ∆2x and is therefore
acceptable for small perturbations. For a function of n variables, (7.6) can be
modified to
@f @f 2
.
.
.
∆x1 1 1
∆xn 1 Oð:∆x: Þ
f ðxÞ 5 f ðx0 Þ 1
@x1 x0
@xn x0
(7.7)
where x0 is the constant vector
x0 5 x10
x20
...
xn0 T
and ∆x denotes the perturbation vector
∆x 5 x1 2x10
x2 2x20
. . . xn 2xn0 T
The term jj∆xjj2 denotes the sum of squares of the entries of the vector (i.e., its
2-norm), which is a measure of the length or “size” of the perturbation vector.1 The
error term dependent on this perturbation is assumed to be small and is neglected in
the sequel.
For nonlinear statespace equations of the form (7.4), let the ith entry of the
vector f be fi. Then applying (7.7) to fi yields the approximation
@fi @fi .
.
.
∆x
1
1
∆xn
1
@x1 ðx0 ; u0 Þ
@xn ðx0 ; u0 Þ
@fi @fi .
.
.
1
∆u
1
1
∆um
1
@u1 ðx0 ; u0 Þ
@ um ðx0 ; u0 Þ
fi ðx; uÞ 5 fi ðx0 ; u0 Þ 1
1
Other norms can be used, but the 2-norm is used most commonly.
(7.8)
243
244
CHAPTER 7 StateSpace Representation
which can be rewritten as
2
∆x1
3
6
7
6 ∆x2 7
6
7
6
7
@fi 6 ^ 7
?
6
7
@xn ðx0 ; u0 Þ 6
7
6 ∆xn21 7
4
5
∆xn
2
3
∆u1
6
7
6 ∆u2 7
6
7
6
7
@fi @fi 6
?
^ 7
1
7
@u1 ðx0 ; u0 Þ
@um ðx0 ; u0 Þ 6
6
7
6 ∆um21 7
4
5
∆um
@fi fi ðx; uÞ 2 fi ðx0 ; u0 Þ 5
@x1 ðx0 ; u0 Þ
(7.9)
In most situations where we seek a linearized model, the nominal state is an
equilibrium point. This term refers to an initial state where the system remains
unless perturbed. In other words, it is a system where the state’s rate of change,
as expressed by the RHS of the state equation, must be zero (see Definition 8.1).
Thus, if the perturbation is about an equilibrium point, then the derivative of the
state vector is zero at the nominal state; that is, fi(x0, u0) is zero and f(x0, u0) is a
zero vector.
The ith entry gi of the vector g can be similarly expanded to yield the perturbation in the ith output
∆yi 5 gi ðx; uÞ 2 gi ðx0 ; u0 Þ
2
3
∆x1
"
#6
7
6 ∆x2 7
@gi @gi 6
?
5
^ 7
7
@x1 ðx0 ; u0 Þ
@xn ðx0 ; u0 Þ 6
4 ∆xn21 5
∆x
2 n
3
∆u1
"
#6
7
6 ∆u2 7
@gi @gi 6
?
^ 7
1
7
@u1 ðx0 ; u0 Þ
@um ðx0 ; u0 Þ 6
4 ∆um21 5
∆um
(7.10)
We also note that the derivative of the perturbation vector is
∆_x 5
d∆x dðx 2 x0 Þ
5
5 x_
dt
dt
because the nominal state x0 is constant.
(7.11)
7.3 Linearization of nonlinear state equations
We now substitute the approximations (7.9) and (7.10) in the state and output
equations, respectively, to obtain the linearized equations
32 ∆x1 3 2
@f1 @f1 7
6
6 @u1 ðx ; u Þ
@xn ðx0 ; u0 Þ 7
∆x2 7
0
0
76
7 6
76
7 6
76
716
? ?
^
?
76
6
7 6
76
7 6
76
7 4 @fn @fn 56
∆x
n21
5
4
?
@xn ðx0 ; u0 Þ
@u1 ðx0 ; u0 Þ
ðx0 ; u0 Þ
∆xn
3 2 ∆x1 3 2
@g1 @g1 @g1 ?
7 6 @u 6
@x1 ðx0 ; u0 Þ
@xn ðx0 ; u0 Þ 7
1 ðx0 ; u0 Þ
∆x2 7
76
7 6
76
7 6
6
76
6
7
? ? ?
?
7:6 ^ 7 1 6
76
7 6
76
6 @g 7
@gn @gn 5 4 ∆xn21 5 4 n ?
@x1 ðx0 ; u0 Þ
@xn ðx0 ; u0 Þ
@u1 ðx0 ; u0 Þ
∆xn
@f1 6 @x1 ðx ; u Þ
0
0
6
6
6
∆_x 5 6
?
6
6 @f 4 n
@x1 2
2
6
6
6
6
∆y 5 6
6
6
4
?
?
?
?
?
?
?
32 ∆u1 3
@f1 7
6
@um ðx0 ; u0 Þ 7
∆u2 7
76
7
76
7
76
?
^ 7
76
7
6
76
7
76
@fn 54 ∆um21 7
5
@um ðx0 ; u0 Þ
∆um
32 ∆u1 3
@g1 7
6
@um ðx0 ; u0 Þ 7
∆u2 7
76
7
76
7
6
76
?
76 ^ 7
7
7
7
76
7
@gn 56
4 ∆um21 5
@um ðx0 ; u0 Þ
∆um
(7.12)
Dropping the ∆s reduces (7.12) to (7.13), with the matrices of the linear state
equations defined as the Jacobians:
@f1 ?
6 @x1 ðx ; u Þ
0
0
6
6
? ?
A56
6
6 @f 4 n
?
@x1 ðx0 ; u0 Þ
2
@g1 ?
6 @x1 ðx ; u Þ
0
0
6
6
C56
? ?
6
6 @g 4 n
?
@x 2
1 ðx0 ; u0 Þ
3
2
@f1 @f1 6 @u1 ðx ; u Þ
@xn ðx0 ; u0 Þ 7
0
0
7
6
6
7
7
6
?
?
7 B56
6 @f 7
@fn 5
4 n
@xn ðx0 ; u0 Þ
@u1 ðx0 ; u0 Þ
3
2
@g1 @g1 6 @u1 ðx ; u Þ
@xn ðx0 ; u0 Þ 7
0
0
7
6
7
6
7 D56
?
?
7
6
7
6 @g @gn 5
4 n
@xn ðx0 ; u0 Þ
@u1 ðx0 ; u0 Þ
?
?
?
?
?
?
3
@f1 @um ðx0 ; u0 Þ 7
7
7
7
?
7
7
@fn 5
@um ðx0 ; u0 Þ
3
@g1 @um ðx0 ; u0 Þ 7
7
7
7
?
7
7
@gn 5
@u
m ðx0 ; u0 Þ
(7.13)
EXAMPLE 7.6
Consider the equation of motion of the nonlinear spring-mass-damper system given by
m€y 1 bðyÞy_ 1 kðyÞ 5 f
where y is the displacement, f is the applied force, m is a mass of 1 kg, b(y) is a nonlinear
damper constant, and k(y) is a nonlinear spring force. Find the equilibrium position corresponding to a force f0 in terms of the spring force, then linearize the equation of motion
about this equilibrium.
245
246
CHAPTER 7 StateSpace Representation
Solution
The equilibrium of the system with a force f0 is obtained by setting all the time derivatives
equal to zero and solving for y to obtain
y0 5 k21 ðf0 Þ
21
where k (•) denotes the inverse function. The equilibrium is therefore at zero velocity and
the position y0.
_ T is
The nonlinear state equation for the system with state vector x5½x1 ; x2 T 5½y; y
x_1
x2
0
5
1
u
x_2
2kðx1 Þ 2 bðx1 Þx2
1
where u 5 f. Then linearizing about the equilibrium, we obtain
2
3
0 1
x_1
0
6 dkðx1 Þ 7 x1
5 42
1
u
5
2bðx
Þj
1
y
0
x_2
x2
1
dx 1
y0
Clearly, the entries of the state matrix are constants whose values depend on the equilibrium position. In addition, terms that are originally linear do not change because of
linearization.
7.4 The solution of linear statespace equations
The statespace equations (7.3) are linear and can therefore be Laplace transformed to obtain their solution. Clearly, once the state equation is solved for the
state vector x, substitution in the output equation easily yields the output vector y.
So we begin by examining the Laplace transform of the state equation.
The state equation involves the derivative x_ of the state vector x. Because
Laplace transformation is simply multiplication by a scalar followed by integration,
the Laplace transform of this derivative is the vector of Laplace transforms of its
entries. More specifically,
L f x_ ðtÞg5 ½sXi ðsÞ 2 xi ð0Þ 5 s½Xi ðsÞ 2 ½xi ð0Þ
5 sXðsÞ 2 xð0Þ
(7.14)
Using a similar argument, the Laplace transform of the product Ax is
# "
#
("
#) "
n
n
n
X
X
X
L fAxðtÞg 5 L
aij xj
aij L fxj g 5
aij Xj ðsÞ
5
j51
j51
j51
5 AXðsÞ
Hence, the state equation
x_ ðtÞ 5 AxðtÞ 1 BuðtÞ;
xð0Þ
(7.15)
7.4 The solution of linear statespace equations
has the Laplace transform
sXðsÞ 2 xð0Þ 5 AxðsÞ 1 BUðsÞ
(7.16)
Rearranging terms, we obtain
½sIn 2 AXðsÞ 5 xð0Þ 1 BUðsÞ
(7.17)
Then premultiplying by the inverse of [sIn 2 A] gives
XðsÞ 5 ½sIn 2A21 ½xð0Þ 1 BUðsÞ
(7.18)
We now need to inverse Laplace transform to obtain the solution of the state
equation. So we first examine the inverse known as the resolvent matrix
1
1 21
(7.19)
In 2 A
½sIn 2A21 5
s
s
This can be expanded as
1
1
In 2 A
s
s
21
5
1
1
1
1
I n 1 A 1 2 A2 1 . . . 1 i Ai 1 . . .
s
s
s
s
(7.20)
Then inverse Laplace transforming yields the series
L 21 f½sIn 2 A21 g 5 In 1 At 1
ðAtÞ2 . . . ðAtÞi . . .
1 1
1
2!
i!
(7.21)
This summation is a matrix version of the exponential function
eat 5 1 1 at 1
ðatÞ2 . . . ðatÞi . . .
1 1
1
2!
i!
(7.22)
It is therefore known as the matrix exponential and is written as
eAt 5
N
X
ðAtÞi
i50
i!
5 L 21 ½sIn 2 A21
(7.23)
Returning to (7.18), we see that the first term can now be easily inversetransformed using (7.23). The second term requires the use of the convolution
property of Laplace transforms, which states that multiplication of Laplace transforms is equivalent to convolution of their inverses. Hence, the solution of the
state equation is given by
ðt
(7.24)
xðtÞ 5 eAt xð0Þ 1 eAðt2τÞ BuðτÞdτ
0
247
248
CHAPTER 7 StateSpace Representation
The solution for nonzero initial time is obtained by simply shifting the time variable to get
xðtÞ 5 eAðt2t0 Þ xð0Þ 1
ðt
eAðt2τÞ BuðτÞdτ
(7.25)
0
The solution includes two terms. The first is due to the initial conditions with
zero input and is known as the zero-input response. The second term is due to the
input with zero initial conditions and is known as the zero-state response. By
superposition, the total response of the system is the sum of the zero-state response
and the zero-input response.
The zero-input response involves the change of the system state from the initial
vector x(0) to the vector x(t) through multiplication by the matrix exponential.
Hence, the matrix exponential is also called the state-transition matrix. This name
is also given to a matrix that serves a similar function in the case of time-varying
systems and depends on the initial as well as the final time and not just the difference between them. However, the matrix exponential form of the state-transition
matrix is only valid for linear time-invariant systems.
To obtain the output of the system, we substitute (7.24) into the output
equation
yðtÞ 5 CxðtÞ 1 DuðtÞ
This gives the time response
ðt
yðtÞ 5 C eAðt2t0 Þ xðt0 Þ 1 eAðt2τÞ BuðτÞdτ 1 DuðtÞ
(7.26)
t0
EXAMPLE 7.7
The state equations of an armature-controlled DC motor are given by
2
3 2
x_1
0
4 x_2 5 5 4 0
x_3
0
1
0
210
32 3 2 3
0
x1
0
1 54 x2 5 1 4 0 5u
x3
211
10
where x1 is the angular position, x2 is the angular velocity, and x3 is the armature current.
Find the following:
1. The state-transition matrix
2. The response due to an initial current of 10 mA with zero angular position and zero
angular velocity
3. The response due to a unit step input
4. The response due to the initial condition in part 2 together with the input in part 3
Solution
1. The state-transition matrix is the matrix exponential given by the inverse Laplace transform
of the matrix
7.4 The solution of linear statespace equations
2
s
6
½sI3 2A21 5 4 0
21
s
0
321
7
21 5
s111
0 10
2
ðs 1 1Þðs 1 10Þ
6
0
4
s 1 11
sðs 1 11Þ
0
210s
sðs 1 1Þðs 1 10Þ
5
1
3
7
s5
2
s
3
s 1 11
1
sðs 1 1Þðs 1 10Þ sðs 1 1Þðs 1 10Þ 7
7
7
7
s 1 11
1
7
7
ðs 1 1Þðs 1 10Þ ðs 1 1Þðs 1 10Þ 7
7
7
7
210
s
5
ðs 1 1Þðs 1 10Þ ðs 1 1Þðs 1 10Þ
3
2
1 1 99
100
1
1 9
10
1
2
1
2
1
6 s 90 s
s 1 1 s 1 10
90 s
s 1 1 s 1 10 7
7
6
6
7
7
6
1
10
1
1
1
1
7
6
2
2
0
7
56
7
6
9 s 1 1 s 1 10
9 s 1 1 s 1 10
7
6
7
6
7
6
10
1
1
1
10
1
5
40
2
2
2
9 s 1 1 s 1 10
9 s 1 10 s 1 1
2
1
6s
6
6
6
60
56
6
6
6
6
40
The preceding operations involve writing s in the diagonal entries of a matrix, subtracting entries of the matrix A, then inverting the resulting matrix. The inversion is feasible
in this example but becomes progressively more difficult as the order of the system
increases. The inverse matrix is obtained by dividing the adjoint matrix by the determinant because numerical matrix inversion algorithms cannot be used in the presence of
the complex variable s.
Next, we inverse Laplace transform to obtain the state-transition matrix
2
61
6
6
6
6
At
e 560
6
6
6
40
3
1
1
ð99 2 100e2t 1 e210t Þ
ð9 2 10e2t 1 e210t Þ 7
90
90
7
7
7
1
1 2t
2t
210t
210t
7
Þ
Þ
ð10e 2 e
ðe 2 e
7
9
9
7
7
7
10 2t
1
210t
210t
2t
Þ
2e Þ 5
ð10e
2 ðe 2 e
9
9
The state-transition matrix can be decomposed as
2
3
0
10 11 1 0
e
e 54 0 0 05 140
10
0
0 0 0
2
At
210
10
210
3
2
21 2t
0
e
1 5
140
9
21
0
1
210
100
3
1
e210t
210 5
90
100
This last form reveals that a matrix exponential is nothing more than a matrix-weighted
sum of scalar exponentials. The scalar exponentials involve the eigenvalues of the state
matrix {0, 21, 210} and are known as the modes of the system. The matrix weights
have rank 1. This general property can be used to check the validity of the matrix
exponential.
249
250
CHAPTER 7 StateSpace Representation
2. For an initial current of 10 mA and zero initial angular position and velocity, the initial
state is
xð0Þ 5 0 0
0:01
T
The zero-input response is
xZI ðtÞ 5 eAt xð0Þ
82
2
3
0
>
< 10 11 1 e0
6
6
7
5 4 0 0 05 140
>
10
:
0
0 0 0
210
10
210
21
3
2
0
2t
7e
6
1 5
140
9
21
0
3
2
9 0
7
1
>6
210t =6 0 7
7
7e
6
210 5
7
6
90 >
1 7
;6
5
4
100
100
3
1
210
100
2 3
2
2
3
3
1
21
1
0
2t
210t
6 7 e
6
7e
6
7e
5405
14 1 5
1 4 210 5
1000
900
9000
0
21
100
By virtue of the decomposition of the state-transition matrix in part 1, the result is
obtained using multiplication by constant matrices rather than ones with exponential
entries.
3. The response due to a step input is easily evaluated starting in the s-domain to avoid
the convolution integral. To simplify the matrix operations, the resolvent matrix is
decomposed as
2
10 11 1
3
2
0
210
21
3
7 1
7
6
1
6
07
10
1 7
5 10s 1 4 0
5 9ðs 1 1Þ
0 0 0
0 210 21
3
2
0
1
1
7
6
1
6
1 4 0 210 210 7
5 90ðs 1 10Þ
6
½sI3 2A21 5 6
4 0
0
0
100
100
The Laplace transform of the zero-state response is
XZS ðsÞ 5 ½sI3 2A21 BUðsÞ
82
3
2
3
0 210 21
< 10 11 1 1
1
5 4 0 0 05
140
10
1 5
:
10s
9ðs 1 1Þ
0 0 0
0 210 21
92 3
2
3
0
1
1
= 0 1
1
4
5
4 0 5
1 0 210 210
90ðs 1 10Þ ;
s
0 100
100
10
2
3
2
3
2 3
1
21
1
10=9
1=9
1 4
4
5
4
5
1 210 5
1
5 0 21
sðs 1 1Þ
sðs 1 10Þ
s
100
21
0
2 3
2
3
2
3 2
3
2
3
1
21
1
1 4
1
1 5 4
1
1 5
4
5
5
4
5
4
5 0 21 1
10=9
2
2
1 210 1=90
s
s
s11
s
s 1 10
0
21
100
7.4 The solution of linear statespace equations
Inverse Laplace transforming, we obtain the solution
3
2
3
2 3 2
21
1
1
7
6
7
6 7 6
xZS ðtÞ 5 4 0 5t 1 4 1 5 10=9 ½1 2 e2t 1 4 210 5 1=90 1 2 e210t
21
100
0
3
2 3 2
2
3
3
1
1
21
211
7
6 7 6
6
7
6
7
2t
5 4 10 5 1=10 1 4 0 5t 1 4 21 5 10=9 e 1 4 10 5 1=90 e210t
2
0
0
2100
1
4. The complete solution due to the initial conditions of part 2 and the unit step input of
part 3 is simply the sum of the two responses obtained earlier. Hence,
xðtÞ 5 xZI ðtÞ 1 xZS ðtÞ
3
3
3
2
2
2 3
2
1
211
1
21
7 e210t 6
7
7 e2t
6
6 7 1
6
7
7
7
7
6
6
6
56
4 0 5 1000 1 4 1 5 900 1 4 210 5 9000 1 4 10 5 1=10
100
0
0
21
3
2 3 2
3
2
1
1
21
7
6 7 6
7
6
2t
210t
7
7 6
7
6
16
4 0 5t 1 4 21 5 10=9 e 1 4 10 5 1=90 e
1
0
2100
3
3 2
3
2
2
2 3
1
21:099
21
1
7
7 6
7
6
6
6 7
22 210t
2t
7
7
6
7
6
6
6
54
1
1407
5 1 4 21 51:11e 1 4 10 51:1 3 10 e
5t
0
2100
1
0
7.4.1 The Leverrier algorithm
The calculation of the resolvent matrix [sI 2 A]21 is clearly the bottleneck in the
solution of statespace equations by Laplace transformation. The Leverrier
algorithm is a convenient method to perform this calculation, which can be programmed on the current generation of handheld calculators that are capable of
performing matrix arithmetic. The derivation of the algorithm is left as an exercise
(see Problem 7.8).
We first write the resolvent matrix as
½sIn 2A21 5
adj½sIn 2 A
det½sIn 2 A
(7.27)
where adj[ ] denotes the adjoint matrix and det[ ] denotes the determinant.
Then we observe that, because its entries are determinants of matrices of order
n 2 1, the highest power of s in the adjoint matrix is n 2 1. Terms of the same
powers of s can be separated with matrices of their coefficients, and the resolvent
matrix can be expanded as
½sIn 2A21 5
P0 1 P1 s 1 . . . 1 Pn21 sn21
a0 1 a1 s 1 . . . 1 an21 sn21 1 sn
(7.28)
251
252
CHAPTER 7 StateSpace Representation
where Pi, i 5 1, 2, . . . , n 2 1 are n 3 n constant matrices and ai, i 5 1, 2, . . . ,
n 2 1 are constant coefficients. The coefficients and matrices are calculated as
follows.
Leverrier algorithm
1. Initialization: k 5 n 2 1
Pn21 5 In
an21 5 2trfAg 5 2
n
X
aii
i51
where tr{ } denotes the trace of the matrix.
2. Backward iteration: k 5 n 2 2, . . . , 0
Pk 5 Pk11 A 1 ak11 In
ak 5 2
1
trf Pk Ag
n2k
3. Check:
½ 0 5 P 0 A 1 a0 I n
The algorithm requires matrix multiplication, matrix scalar multiplication,
matrix addition, and trace evaluation. These are relatively simple operations
available in many handheld calculators, with the exception of the trace operation.
However, the trace operation can be easily programmed using a single repetition
loop. The initialization of the algorithm is simple, and the backward iteration
starts with the formulas
1
an22 5 2 trf Pn22 Ag
2
Pn22 5 A 1 an21 In
The Leverrier algorithm yields a form of the resolvent matrix that cannot be
inverse Laplace transformed directly to obtain the matrix exponential. It is first
necessary to expand the following s-domain functions into the partial fractions:
sn21
n
L ðs 2 λi Þ
5
n
X
qi;n21
i51
s 2 λi
i51
;
sn22
n
L ðs 2 λi Þ
i51
5
n
X
qi;n22
i51
s 2 λi
;...;
1
n
L ðs 2 λi Þ
n
X
qi;0
5
s
2
λi
i51
(7.29)
i51
where λi, i 5 1, . . . , n are the eigenvalues of the state matrix A defined by
det½sIn 2 A 5 sn 1 an21 sn21 1 . . . 1 a1 s 1 a0 5 ðs 2 λ1 Þðs 2 λ2 Þ. . .ðs 2 λn Þ (7.30)
To simplify the analysis, we assume that the eigenvalues in (7.30) are distinct
(i.e., λi 6¼λj, i6¼j). The repeated eigenvalue case can be handled similarly but
requires higher-order terms in the partial fraction expansion.
7.4 The solution of linear statespace equations
Substituting in (7.28) and combining similar terms gives
½sIn 2A21 5 Pn21
n
X
qi;n21
i51
5
5
1
s 2 λ1
n
X
i51
1 . . . 1 P1
s 2 λi
n21
X
1
q1; j Pj 1
i50
"
1
s 2λi
n
n
X
X
qi;1
qi;0
1 P0
s2λ
s
2 λi
i
i51
i51
n21
X
1
s 2 λ2
#
n21
X
q2; j Pj 1 . . . 1
i50
n21
1 X
qn; j Pj
s 2 λn i50
qi; j Pj
i50
Thus, we can write the resolvent matrix in the simple form
½sIn 2A21 5
n
X
i51
1
Zi
s 2 λi
where the matrices Zi, i 5 1, 2, . . . , n are given by
"
#
n21
X
Zi 5
qi; j Pj
(7.31)
(7.32)
j50
Finally, we inverse the Laplace transform to obtain the matrix exponential
eAt 5
n
X
Zi eλi t
(7.33)
i51
This is the general form of the expansion used to simplify the computation in
Example 7.7 and is the form we use throughout this text.
EXAMPLE 7.8
Use the Leverrier algorithm to compute the matrix exponential for the state matrix of
Example 7.7:
2
3
0
1
0
A540
0
1 5
0 210 211
Solution
1. Initialization:
P2 5 I3
a2 5 2 trfAg 5 11
2. Backward iteration: k 5 1, 0
(a) k 5 1
2
11
P1 5 A 1 a2 I3 5 A 1 11 I3 5 4 0
0
3
1
0
11
15
210 0
253
254
CHAPTER 7 StateSpace Representation
82
>
< 11
1
1
a1 5 2 trfP1 Ag 5 2 tr 6
4 0
2
2 >
:
0
82
> 0
1 <6
5 2 tr 4 0
2 >
:
0
11
210
0
1
11
0
32
0
1
0
210
76
1 54 0
210 0
0
39
>
=
7
1 5
>
;
211
0
39
1
>
=
7
0 5 5 10
>
;
210
(b) k 5 0
2
11
P0 5 P1 A 1 a1 I3 5 4 0
0
32
1
0
0
11
15 40
210 0
0
3 2
3 2
3
0
10 0 0
10 11 1
1 5 1 4 0 10 0 5 5 4 0 0 0 5
211
0 0 10
0 0 0
1
0
210
82
32
10 11 1
0
1
1 <4
a0 5 2 trfP0 Ag 5 2 tr
0 0 0 54 0
3
3 :
0 0 0
0
1
0
210
82
39
0
=
< 0
1
1 5 5 2 tr 4 0
;
3 :
211
0
3. Check:
2
0
½0 5 P0 A 1 a0 In 5 5 4 0
0
1
0
210
32
3
0
10 11 1
1 54 0 0 0 5 5 ½0
211
0 0 0
Thus, we have
2
3 2
3
2
10 11 1
11
1
0
1 0
4 0 0 0514 0
11
1 5s 1 4 0 1
0 0 0
0 210 0
0 0
½sIn 2A21 5
10s 1 11s2 1 s3
The characteristic polynomial of the system is
s3 1 11s2 1 10s 5 sðs 1 1Þðs 1 10Þ
and the system eigenvalues are {0, 21, 210}.
Next, we obtain the partial fraction expansions
s2
3
5
L ðs 2 λi Þ
3
X
21=9
10=9
qi;2
0
1
5 1
s
1
1
s
1 10
s
2
λ
s
i
i51
i51
s
3
L ðs 2 λi Þ
i51
5
3
X
1=9
21=9
qi;1
0
1
5 1
s11
s 1 10
s 2 λi
s
i51
3
0
0 5s2
1
39
0 0 =
0 05 50
;
0 0
7.4 The solution of linear statespace equations
1
5
3
L ðs 2 λi Þ
3
X
1=10
21=9
1=90
qi;0
1
1
5
s
s11
s 1 10
s 2 λi
i51
i51
where some of the coefficients are zero due to cancellation.
the matrices
2
3
10 11 1
4
0 0 05
Z1 5 1=10
0 0 0
2
3
2
3
2
10 11 1
11 1 0
1 0
4
5
4
5
Z2 5 21=9 0 0 0 1 1=9 0 11 1 1 21=9 4 0 1
0 0 0
0 210 0
0 0
2
3
2
3
2
10 11 1
11 1 0
1 0
6
7
6
7
6
Z35 1=90 4 0 0 0 5 1 21=9 4 0 11 1 5 1 10=9 4 0 1
2
0 0 0
0
1
1
0 210 0
3
This allows us to evaluate
2
3
3
0
0 210 21
5
4
0 5 1=9 0 10 1 5
1
0 210 21
3
0
7
05
0 0 1
6
7
5 1=90 4 0 210 210 5
0 100 100
Therefore, the state-transition matrix is
2
3
0
10 11 1
0
e
eAt 5 4 0 0 0 5 1 4 0
10
0
0 0 0
2
210
10
210
3
2
21 2t
0
e
1 5
140
9
21
0
1
210
100
3
1
e210t
210 5
90
100
Thus, we obtained the answer of Example 7.7 with far fewer partial fraction expansions
but with some additional operations with constant matrices. Because these operations
are easily performed using a calculator, this is a small price to pay for the resulting
simplification.
7.4.2 Sylvester’s expansion
The matrices Zi, i 5 1, 2, . . . , n, obtained in Section 7.4.1 using the Leverrier
algorithm, are known as the constituent matrices of A. The constituent matrices
can also be calculated using Sylvester’s formula as follows:
n
L ½A 2 λj In Zi 5
j51
j 6¼ i
n
L ½λi 2 λj j51
j 6¼ i
where λi, i 5 1, 2, . . . , n are the eigenvalues of the matrix A.
(7.34)
255
256
CHAPTER 7 StateSpace Representation
Numerical computation of the matrix exponential using (7.34) can be problematic. For example, if two eigenvalues are almost equal, the scalar denominator in
the equation is small, resulting in large computational errors. In fact, the numerical computation of the matrix exponential is not as simple as our presentation
may suggest, with all known computational procedures failing in special cases.
These issues are discussed in more detail in a well-known paper by Moler and
Van Loan (1978).
EXAMPLE 7.9
Calculate constituent matrices of the matrix A given in Examples 7.7 and 7.8 using
Sylvester’s formula.
Solution
3
L A 2 λj I n
Z1 5
j51
j 6¼ i
3
L 0 2 λj
2
3
10 11 1
ðA 1 In ÞðA 1 10In Þ
5
5 1=10 4 0 0 0 5
ð1Þð10Þ
0 0 0
j51
j 6¼ i
3
L A 2 λj In
Z2 5
j51
j 6¼ i
3
L 21 2 λj
2
0
AðA 1 10In Þ
5 1=9 4 0
5
ð 21Þð 21 1 10Þ
0
210
10
210
3
21
1 5
21
j51
j 6¼ i
3
L ½A 2 λj In 2
0
AðA 1 In Þ
Z3 5 3
5
5 1=90 4 0
ð2 10Þð2 10 1 1Þ
0
L ½ 210 2 λj j51
j 6¼ i
1
210
100
3
1
210 5
100
j51
j 6¼ i
The constituent matrices can be used to define any analytic function of a
matrix (i.e., a function possessing a Taylor series) and not just the matrix exponential. The following identity is true for any analytic function f(λ):
f ðAÞ 5
n
X
Zi f ðλi Þ
(7.35)
i51
This identity allows us to calculate the functions eAt and Ai, among others. Using
the matrices Zi described in Example 7.9 and (7.35), we obtain the state-transition
matrix described in Example 7.7.
7.4 The solution of linear statespace equations
7.4.3 The state-transition matrix for a diagonal state matrix
For a state matrix Λ in the form
Λ 5 diagfλ1 ; λ2 ; . . .; λn g
the resolvent matrix is
21
½sIn 2Λ
(7.36)
1
1
1
5 diag
;
; . . .;
s 2 λ1 s 2 λ2
s 2 λn
(7.37)
The corresponding state-transition matrix is
eΛt 5 L f½sIn 2 Λ21 g
5 diagfeλ1 t ; eλ2 t ; . . .; eλn t g
5 diagf1; 0; . . .; 0geλ1 t 1 diagf0; 1; . . .; 0geλ2 t 1 . . . 1 diagf0; 0; . . .; ngeλn t
(7.38)
Thus, the ith constituent matrix for the diagonal form is a diagonal matrix with
unity entry (i, i) and all other entries equal to zero.
EXAMPLE 7.10
Calculate the state-transition matrix if the state matrix Λ is
Λ 5 diagf21; 25; 26; 220g
Solution
Using (7.38), we obtain the state-transition matrix
eΛt 5 diagfe2t ; e25t ; e26t ; e220t g
Assuming distinct eigenvalues, the state matrix can in general be written in
the form
A 5 VΛV 21 5 VΛW
(7.39)
where
v2
h
V 5 v1
2
?
i
vn
3
wT1
6 7
6 wT 7
6 2 7
6
7
7
W 56
6
7
6 ^ 7
6
7
4 5
wTn
vi, wi, 5 i 5 1, 2, . . . , n are the right and left eigenvectors of the matrix A, respectively. The fact that W is the inverse of V implies that their product is the identity
matrix—that is,
WV 5 wTi vj 5 In
257
258
CHAPTER 7 StateSpace Representation
Equating matrix entries gives
wTi vj
5
1;
0;
i5j
i 6¼ j
(7.40)
The right eigenvectors are the usual eigenvectors of the matrix, whereas the left
eigenvectors can be shown to be the eigenvectors of the matrix transpose.
The matrix A raised to any integer power i can be written in the form
Ai 5 ðVΛ W ÞðVΛ W Þ . . . ðVΛ W Þ 5 VΛi W
(7.41)
Substituting from (7.41) in the matrix exponential series expansion gives
eAt 5
N
X
ðAtÞi
i50
5
i!
N
X
VðΛ tÞi W
i50
i!
"
#
N
X
diagfðλ1 tÞi ; ðλ2 tÞi ; ?; ðλn tÞi g
5V
W
i!
i50
"
(
N
N
N
X
X
ðλ1 tÞi X
ðλ2 tÞi
ðλn tÞi
;
; ?;
5 V diag
i!
i!
i!
i50
i50
i50
)#
W
That is,
eAt 5 VeΛt W
(7.42)
Thus, we have an expression for the matrix exponential of A in terms of the
matrix exponential for the diagonal matrix Λ. The expression provides another
method to calculate the state-transition matrix using the eigenstructure (eigenvalues and eigenvectors) of the state matrix. The drawback of the method is that the
eigenstructure calculation is computationally costly.
EXAMPLE 7.11
Obtain the constituent matrices of the state matrix using (7.42), then obtain the statetransition matrix for the system
x_ ðtÞ 5
0
22
1
xðtÞ
23
7.4 The solution of linear statespace equations
Solution
The state matrix is in companion form, and its characteristic equation can be directly written
with the coefficient obtained by negating the last row of the matrix
λ2 1 3λ 1 2 5 ðλ 1 1Þðλ 1 2Þ 5 0
The system eigenvalues are therefore { 2 1, 22}. The matrix of eigenvectors is the Van der
Monde matrix
22 21
1
1
2
1
1
1
5
W 5 V 21 5
V5
21 21
21 22
2211
The matrix exponential is
1
1
3
2
3 e2t
0
"
#
1
7 2
6
At
Λt
7
6
4
5
e 5 Ve W 5
4 5
21 21
21 22
0
e22t
3
2
1
#
("
) 2
"
#
h
i
h
i
1
1
7
6
6 7
5
e2t 0 1
0 e22t
5
4
21
22
21 21
"
#
"
#
1
1
2t
5
2 1 e 1
21 21 e22t
21
22
"
#
"
#
2
1
21 21
5
e2t 1
e22t
22 21
2
2
2
e
0
?
λ2 t
?
^
&
0
2 T 3
3 w1
7
0 6
76
wT2 7
6
7
7
0 7
76
6
7
76
7
7
^ 56 ^ 7
6
7
eλn t 4 5
T
wn
3
eΛt W
e λ1 t
6
6 0
6
56
6 ^
4
0
2
As one may infer from the partitioned matrices used in Example 7.11,
expression (7.42) can be used to write the state-transition matrix in terms of the
constituent matrices. We first obtain the product
?
3
3
2
2
01 3 n
01 3 n
wT1
6 7
6 7
6 7
6 01 3 n 7
6 wT 7
6 01 3 n 7
7
7
6
6 2 7
6
7 λt 6
7 λt
7
6
6
7e 1 1 6 7e 2 1 . . . 1 6 7eλ1 t
56
7
6 ^ 7
6
6 ^ 7
7
7
6
6 ^ 7
6
7
7
6
6
6 7
5
4 5
4 5
4
01 3 n
01 3 n
wTn
2
259
CHAPTER 7 StateSpace Representation
Then we premultiply by the partition matrix V to obtain
82 T 3
2
3
2
3 9
01 3 n
01 3 n
w1
>
>
>
>
>
>
>
6 7
6 7
6 7 >
>
>
>
6 01 3 n 7
6 wT 7
6 01 3 n 7 >
>
>
>
>
6
7
6
7
6
7
2
=
h
i<6
7
6
7
6
7
Λt
λ1 t
λ2 t
λn t
..
.
6
7
6
7
6
7
Ve W 5 v1 v2 ? vn
1
1
1
e
e
e
6 ^ 7
6
7
6 ^ 7 >
>
>
7
6 ^ 7
6
7 >
>6
>
>
6
7
6
7
6
7 >
>
>
>
4
5
4
5
4
5 >
>
>
>
>
;
: 01 3 n
wTn
01 3 n
260
(7.43)
Substituting from (7.43) into (7.42) yields
eAt 5
n
X
Zi eλi t 5
n
X
i51
vi wTi eλi t
(7.44)
i51
Hence, the ith constituent matrix of A is given by the product of the ith right and
left eigenvectors. The following properties of the constituent matrix can be
proved using (7.44). The proofs are left as an exercise.
Properties of constituent matrices
1. Constituent matrices have rank 1.
2. The product of two constituent matrices is
Zi ; i 5 j
Zi Z j 5
0; i 6¼ j
Raising Zi to any power gives the matrix Zi. Zi is said to be idempotent.
3. The sum of the n constituent matrices of an n 3 n matrix is equal to the
identity matrix
n
X
Zi 5 In
i51
EXAMPLE 7.12
Obtain the constituent matrices of the state matrix of Example 7.11 using (7.44), and verify that they satisfy Properties 1 through 3. Then obtain the state-transition matrix for the
system.
Solution
The constituent matrices are
Z1 5
Z2 5
1
21
1
22
2
1 5
2
22
21
21 5
21
2
1
21
21
2
7.4 The solution of linear statespace equations
Both matrices have a second column equal to the first and clearly have rank 1. The product of the first and second matrices is
"
Z1 Z2 5
"
5
#"
2
1
22
21
# "
0
0
0
0
5
21
2
21
21
#
2
#"
21
2
1
#
5 Z2 Z1
22
21
2
22
1
5
21
2
22
1
21
21
2
21
5
2
21
2
21
2
2
2
The squares of the matrices are
Z1 Z1 5
2
22
1
21
Z2 Z2 5
21
2
21
2
The state-transition matrix is
eAt 5
2
X
Zi eλi t 5
i51
2
22
1
e2t 1
21
21
2
21 22t
e
2
EXAMPLE 7.13
Obtain the state-transition matrix for the system with state matrix
2
3
0 1 3
4
A5 3 5 65
5 6 7
Solution
Using the MATLAB command eig, we obtain the matrices
2
0:8283
V 5 4 0:1173
20:5479
20:2159
20:6195
20:7547
3
2
0:5013
0:8355
5
20:7931
W 5 4 20:4050
0:3459
0:4399
0:3121
20:5768
20:7641
3
20:4952
20:7357 5
0:5014
and the eigenvalues { 2 1.8429, 13.3554, 0.4875}. Then multiplying each column of V by
the corresponding row of W, we obtain the constituent matrices
2
0:6920
Z1 5 4 0:0980
20:4578
2
0:2205
Z3 5 4 20:3489
20:1521
0:2585
0:0366
20:1710
3
20:4102
20:0581 5
0:2713
20:3831
0:6061
20:2643
3
0:2513
20:3977 5
0:1734
2
3
0:0874 0:1245 0:1588
4
Z2 5 0:2509 0:3573 0:4558 5
0:3057 0:4353 0:5552
261
262
CHAPTER 7 StateSpace Representation
The state-transition matrix is
3
3
2
2
0:6920
0:2585
20:4102
0:0874 0:1245 0:1588
7 21:8429t 6
7 13:3554t
6
7
eAt 5 6
0:0366
20:0581 7
16
5e
4 0:0980
4 0:2509 0:3573 0:4558 5e
20:4578 20:1710
0:2713
0:3057 0:4353 0:5552
3
2
0:2205
20:3831
0:2513
7 0:4875t
6
7e
16
20:3489
0:6061
20:3977
5
4
20:1521 20:2643
0:1734
7.4.4 Real form for complex conjugate eigenvalues
For the case of complex conjugate eigenvalues of a real matrix, the eigenvectors
will also be complex conjugate. This is easily shown by taking the complex conjugate of the equation that defines the right eigenvectors
Av 5 λv 5 Av 5 λv
(7.45)
where we used the fact that the real matrix A is identical to its conjugate.
Similarly, for the left eigenvectors
wT A 5 λwT 5 wT A 5 λwT
(7.46)
Thus, the constituent matrix for a complex eigenvalue λ as given by
Z 5 vwT
(7.47)
is the conjugate of the constituent matrix of its complex conjugate eigenvalue λ
Z 5 vwT
(7.48)
Using this fact, we can simplify the expansion of the state-transition matrix with
complex eigenvalues as follows:
eAt 5 Zeλt 1 Zeλt 5 2RefZeλt g
(7.49)
For the eigenvalue λ 5 σ 1 jωd we have
eAt 5 2RefZgeσT cosðωd tÞ 2 2 ImfZgeσT sinðωd tÞ
(7.50)
EXAMPLE 7.14
The RLC resonance circuit of Figure 7.3 with capacitor voltage as output has the
statespace model
2
3
0
1
1
R
0
x_ 5 4 2
u
2 5x 1
1
LC
L
7.4 The solution of linear statespace equations
"
1
y 5 LC
#
0 x
x 5 x1
x 2 T
1
Assuming normalized component values such that RL 5 2; LC
5 10, obtain a real form of the
state-transition matrix of the system by combining complex conjugate terms.
L
R
C
FIGURE 7.3
Series RLC circuit.
Solution
The state matrix is in companion form, and the characteristic equation can be written by
inspection:
λ2 1
R
1
λ1
5 λ2 1 2λ 1 10 5 0
L
LC
The eigenvalues of the system are
λ1;2 5 2 1 6 j3
The right eigenvector matrix is the Van der Monde matrix
V5
1
21 1 j3
1
21 2 3j
The left eigenvector matrix is
W 5 V 21 5
1 32j
6 31j
2j
j
The constituent matrix for the first eigenvalue is
Z 5 vwT 5
1
1
32j
21 1 j3 6
2j 5
1 32j
6 10j
1 1 0
j
2j
5
1
31j
2 0 1
6
The state transition matrix can be written as
eAt 5
1
1 0 2t
e cosð3tÞ 2
0 1
3
21
10
21 2t
e sinð3tÞ
1
21
10
21
1
263
264
CHAPTER 7 StateSpace Representation
7.5 The transfer function matrix
The transfer function matrix of a system can be derived from its state and output
equations. We begin by Laplace transforming the state equation (7.3) with zero
initial conditions to obtain
XðsÞ 5 ½sIn 2A21 BUðsÞ
(7.51)
Then we Laplace transform the output equation and substitute from (7.51) to get
YðsÞ 5 C½sIn 2A21 BUðsÞ 1 DUðsÞ
The last equation can be rewritten in the form
YðsÞ 5 HðsÞUðsÞ
L
! yðtÞ 5 H ðtÞ uðtÞ
H ðsÞ 5 C½sIn 2A21 B 1 D
L
! H ðtÞ 5 CeAt B 1 DδðtÞ
(7.52)
where H(s) is the transfer function matrix and H(t) is the impulse response matrix.
The equations emphasize the fact that the transfer function and the impulse
response are Laplace transform pairs.
The preceding equation cannot be simplified further in the MIMO case
because division by a vector U(s) is not defined. In the SI case, the transfer function can be expressed as a vector of ratios of outputs to inputs. This reduces to a
single scalar ratio in the SISO case. In general, the ijth entry of the transfer function matrix denotes
Yi ðsÞ (7.53)
hij ðsÞ 5
Uj ðsÞ Ul 50; l6¼j
zero initial conditions
Equation (7.52) can be rewritten in terms of the constituent matrices of A as
"
#
n
X
1
Zi
B1D
HðsÞ 5 C
s 2 λi
i51
Thus,
HðsÞ 5
n
X
i51
CZi B
1
1D
s 2 λi
L
! HðtÞ 5
n
X
CZi Beλit 1 DδðtÞ
(7.54)
i51
This shows that the poles of the transfer function are the eigenvalues of the state
matrix A. In some cases, however, one or both of the matrix products CZi, ZiB are
zero, and the eigenvalues do not appear in the reduced transfer function.
The evaluation of (7.52) requires the computation of the resolvent matrix
and can be performed using the Leverrier algorithm or Sylvester’s formula.
Nevertheless, this entails considerable effort. Therefore, we should only use
7.5 The transfer function matrix
(7.52) to evaluate the transfer function if the statespace equations are given
and the input-output differential equations are not. It is usually simpler to
obtain the transfer function by Laplace transforming the input-output differential equation.
EXAMPLE 7.15
Calculate the transfer function of the system of Example 7.7 with the angular position as
output.
Solution
3
2
82
10 11 1
0 210
>
>
1
>
6
7
6
>
>
0
0
0
1
0
10
4
5
4
>
>
>
< 0 0 0 10s
0 210
2
3
H ðsÞ 5 1 0 0
>
0
1
1
>
>
>
1
6
7
>
>
4 0 210 210 5
>
>
90ðs 1 10Þ
:
0 100
100
1
10
1
1
5 2
s
9ðs 1 1Þ 9ðs 1 10Þ
5
3
9
>
>
1
>
7
>
1 5
1>
>2 3
9ðs 1 1Þ >
>
= 0
21
6 7
4 0 5
>
>
>
10
>
>
>
>
>
;
21
10
sðs 1 1Þðs 1 10Þ
7.5.1 MATLAB commands
MATLAB obtains the transfer function for the matrices (A, B, C, D) with the
commands
.. g 5 tfðssðA; B; C; DÞÞ
For example, the matrices
2
0
A54 0
23
C5
0
0
1
1
24
1 0
1 1
3
0
1 5
22
2
3
0
05
1
0 1
D5
0 1
0
B541
0
are entered as
.. A 5 ½zerosð2; 1Þ; ½1; 0; 1; 1; 23; 24; 22; B 5 ½zerosð1; 2Þ; eyeð2Þ
.. C 5 ½zerosð2; 1Þ; ½1; 0; 1; 1; D 5 ½zerosð2; 1Þ; onesð2; 1Þ
Then the transfer function for the first input is obtained with the transformation
command
.. g 5 tfðssðA;B;C;DÞÞ
265
266
CHAPTER 7 StateSpace Representation
Transfer function from input 1 to output
s^2 1 2s
#1: – – – – – – – – – – –
s^3 1 s^2 1 2s 1 3
s^2 2 2s 2 3
#2: – – – – – – – – – – – –
s^3 1 s^2 1 2s 1 3
Transfer function from input 2 to output
s^3 1 s^2 1 3s 1 3
#1: – – – – – – – – – – –
s^3 1 s^2 1 2s 1 3
s^3 1 2s^2 1 2s 1 3
#2: – – – – – – – – – – – –
s^3 1 s^2 1 2s 1 3
The first two terms are the first column of the transfer function matrix
s2 1 2s
s2 2 2s 2 3
H1 ðsÞ 5 3
s 1 s2 1 2s 1 3
The next two terms are transfer function column corresponding to the second
input.
The tf command can also be used to obtain the resolvent matrix by setting
B 5 C 5 In with D zero. The command takes the form
.. Resolvent 5 tfðssðA; eyeðnÞ; eyeðnÞ; 0ÞÞ
7.6 Discrete-time statespace equations
Given an analog system with piecewise constant inputs over a given sampling
period, the system state variables at the end of each period can be related by a
difference equation. The difference equation is obtained by examining the solution of the analog state derived in Section 7.4, over a sampling period T. The
solution is given by (7.25), which is repeated here for convenience:
ð tf
Aðtf 2t0 Þ
xðtf Þ 5 e
xðt0 Þ 1
eAðtf 2τÞ BuðτÞdτ
(7.55)
t0
Let the initial time t0 5 kT and the final time tf 5 (k 1 1)T. Then the solution (7.55)
reduces to
ð ðk11ÞT
xðk 1 1Þ 5 eAT xðkÞ 1
eAððk11ÞT2τÞ BuðτÞdτ
kT
(7.56)
7.6 Discrete-time statespace equations
where x(k) denotes the vector at time kT. For a piecewise constant input
uðtÞ 5 uðkÞ;
kT # t , ðk 1 1ÞT
the input can be moved outside the integral. The integrand can then be simplified
by changing the variable of integration to
λ 5 ðk 1 1ÞT 2 τ
dλ 5 2dτ
The integral now becomes
ð 0
ð T
eAλ Bð2 dλÞ uðkÞ 5
eAλ Bdλ uðkÞ
T
0
Substituting in (7.56), we obtain the discrete-time state equation
xðk 1 1Þ 5 Ad xðkÞ 1 Bd uðkÞ
(7.57)
Ad 5 eAT
(7.58)
where
Bd 5
ðT
eAλ Bdλ
(7.59)
0
Ad is the discrete-time state matrix and Bd is the discrete input matrix, and they
are clearly of the same orders as their continuous counterparts. The discrete-time
state matrix is the state-transition matrix for the analog system evaluated at the
sampling period T.
Equations (7.58) and (7.59) can be simplified further using properties of the
matrix exponential. For invertible state matrix A, the integral of the matrix exponential is
ð
(7.60)
eAt dt 5 A21 ½eAt 2 In 5 ½eAt 2 In A21
This allows us to write Bd in the form
Bd 5 A21 ½eAT 2 In B 5 ½eAT 2 In A21 B
(7.61)
Using the expansion of the matrix exponential (7.33), we rewrite (7.58) and
(7.59) as
Ad 5
n
X
i51
Zi eλi T
(7.62)
267
268
CHAPTER 7 StateSpace Representation
X
n
λi τ
Bdτ
Zi e
ÐT
Bd 5
0
i51
5
n
X
ðT
Zi B
(7.63)
λi τ
e
dτ
0
i51
The integrands in (7.63) are scalar functions, and the integral can be easily
evaluated. Because we assume distinct eigenvalues, only one eigenvalue can be
zero. Hence, we obtain the following expression for Bd:
2
3
8
n
>
λi T
X
>
1
2
e
>
5
>
λi 6¼ 0
Z i B4
>
>
>
2 λi
< i51
(7.64)
Bd 5
2
3
>
>
n
λi T
X
>
1
2
e
>
>
5 λ1 5 0
>
Z BT 1
Z i B4
>
: 1
2 λi
i52
The output equation evaluated at time kT is
yðkÞ 5 CxðkÞ 1 DuðkÞ
(7.65)
The discrete-time statespace representation is given by (7.57) and (7.65).
Equation (7.57) is approximately valid for a general input vector u(t) provided
that the sampling period T is sufficiently short. The equation can therefore be
used to obtain the solution of the state equation in the general case.
EXAMPLE 7.16
Obtain the discrete-time state equations for the system of Example 7.7
2
3 2
x_1
0
4 x_2 5 5 4 0
x_3
0
1
0
210
32 3 2 3
0
x1
0
1 54 x2 5 1 4 0 5u
x3
211
10
for a sampling period T 5 0.01 s.
Solution
From Example 7.7, the state-transition matrix of the system is
2
3
0
10 11 1
e0 4
4
5
1 0
e 5 0 0 0
10
0
0 0 0
2
At
210
10
210
3
2
21 2t
0
e
5
1
140
9
21
0
1
210
100
3
1
e210t
210 5
90
100
Thus, the discrete-time state matrix is
2
3
0
10 11 1
0
e
54 0 0 05 140
10
0
0 0 0
2
Ad 5 e
0:01 3 A
210
10
210
3
2
21 20:01
0
e
1 5
140
9
21
0
1
210
100
3
1
e210 3 0:01
210 5
90
100
7.6 Discrete-time statespace equations
This simplifies to
2
3
0:1
0:0
0:9995
0:0095 5
20:0947 0:8954
1:0
Ad 5 4 0:0
0:0
The discrete-time input matrix is
Bd 5 Z1 Bð0:01Þ 1 Z2 Bð1 2 e20:01 Þ 1 Z3 Bð1 2 e210 3 0:01 Þ=10
3
3
3 2
2
2
21
1
0:01
7
7
7 6
6
6
20:01
210 3 0:01
7
7
7 6
Þ16
Þ
56
4 210 5 1=90 ð1 2 e
4 0 5 1 4 1 5 10=9 ð1 2 e
21
100
0
This simplifies to
2
3
1:622 3 1026
24
Bd 5 4 4:821 3 10 5
9:468 3 1022
7.6.1 MATLAB commands for discrete-time statespace equations
The MATLAB command to obtain the discrete statespace quadruple pd from
the continuous quadruple p with sampling period T 5 0.1 is
.. pd 5 c2dðp;:1Þ
Alternatively, the matrices are obtained using (7.58) and (7.61) and the MATLAB
commands
.. Ad 5 expmðA 0:1Þ
.. Bd 5 A\ðAdeyeðsizeðAÞÞÞ B
7.6.2 Complex conjugate eigenvalues
If an analog system with complex conjugate eigenvalues λ1,2 5 σ 6 jωd is discretized with the analog input constant over each sampling period, then the resulting
system has the complex conjugate eigenvalues eλ1;2 T 5 eσ 6 jωd . From (7.50), the
constituent matrices for the discrete state matrix are the same as those for the
analog system, and the discrete state matrix is in the form
eAT 5 2RefZgeσT cosðωd TÞ 2 2ImfZgeσT sinðωd TÞ
The discrete input matrix is in the form
!
λT
ðT
e 21
eλT 2 1
Aτ
Bd 5
1 ZB
e Bdτ 5 ZB
λ
λ
0
λT
λT
e 21
e 21
5 2 RefZgBRe
2 2 ImfZgBIm
λ
λ
(7.66)
269
270
CHAPTER 7 StateSpace Representation
eλT 2 1 σ½eσT cosðωd TÞ 2 1 1 ωd eσT sinðωd TÞ
5
λ
jλj2
2j
ωd ½eσT cosðωd TÞ 2 1 2 σeσT sinðωd TÞ
jλj2
σ½eσT cosðωd TÞ 2 1 1 ωd eσT sinðωd TÞ
Bd 5 2RefZgB
jλj2
12ImfZgB
ωd ½eσT cosðωd TÞ 2 1 2 σeσT sinðωd TÞ
jλj2
(7.67)
For eigenvalues on the imaginary axis λ 5 jωd, and we have the simpler forms
eAT 5 2RefZg cosðωd TÞ 2 2ImfZg sinðωd TÞ
Bd 5 2RefZgB
sinðωd TÞ
cosðωd TÞ 2 1
1 2ImfZgB
ωd
ωd
(7.68)
(7.69)
EXAMPLE 7.17
Obtain the discrete state and input matrix for the series resonances circuit of Example 7.14
with a DAC and ADC and a sampling period T. Evaluate the matrices for T 5 0.05 s and
check your answer with MATLAB.
Solution
The state matrix is
1 21 21 2T
1 0 2T
e sinð3TÞ
e cosð3TÞ 2
1
0 1
3 10
1 1 0
j 21 21
j 21
0
0
1
5
1
2ZB 5 2
1
1
1
2 0 1
6 10
3 1
Ad 5 eAT 5
Bd 5
0
1
2 ½e2T cosð3TÞ 2 1 1 3e2T sinð3TÞ 1
1
10
3
21 3½e2T cosð3TÞ 2 1 1 e2T sinð3TÞ
1
10
Evaluating the matrices at T 5 0.05 s gives the same answer as the MATLAB commands
.. expmðA :05Þ
ans 5
0:9879 0:0474
2 0:4738 0:8932
.. Bd 5 A\ðAd 2 eyeðsizeðAÞÞÞ B
Bd 5
0:0012
0:0474
7.7 Solution of discrete-time statespace equations
7.7 Solution of discrete-time statespace equations
We now seek an expression for the state at time k in terms of the initial condition
vector x(0) and the input sequence u(k), k 5 0, 1, 2, . . ., k 2 1. We begin by examining the discrete-time state equation (7.57):
xðk 1 1Þ 5 Ad xðkÞ 1 Bd uðkÞ
At k 5 0, 1, we have
xð1Þ 5 Ad xð0Þ 1 Bd uð0Þ
xð2Þ 5 Ad xð1Þ 1 Bd uð1Þ
(7.70)
Substituting from the first into the second equation in (7.70) gives
xð2Þ 5 Ad ½Ad xð0Þ 1 Bd uð0Þ 1 Bd uð1Þ
5 A2d xð0Þ 1 Ad Bd uð0Þ 1 Bd uð1Þ
(7.71)
We then rewrite (7.71) as
xð2Þ 5 A2d xð0Þ 1
221
X
A22i21
Bd uðiÞ
d
(7.72)
i50
and observe that the expression generalizes to
xðkÞ 5 Akd xð0Þ 1
k21
X
Adk2i21 Bd uðiÞ
(7.73)
i50
This expression is in fact the general solution. Left as an exercise are details of
the proof by induction where (7.73) is assumed to hold and it is shown that a
similar form holds for x(k 1 1).
The expression (7.73) is the solution of the discrete-time state equation. The
matrix Akd is known as the state-transition matrix for the discrete-time system,
and it plays a role analogous to its continuous counterpart. A state-transition matrix
can be defined for time-varying discrete-time systems, but it is not a matrix power,
and it is dependent on both time k and initial time k0.
The solution (7.73) includes two terms as in the continuous-time case. The first
is the zero-input response due to nonzero initial conditions and zero input. The
second is the zero-state response due to nonzero input and zero initial conditions.
Because the system is linear, each term can be computed separately and then added
to obtain the total response for a forced system with nonzero initial conditions.
Substituting from (7.73) in the discrete-time output equation (7.65) gives the
output
(
)
k21
X
Ak2i21
Bd uðiÞ 1 DuðkÞ
(7.74)
yðkÞ 5 C Akd xð0Þ 1
d
i50
271
272
CHAPTER 7 StateSpace Representation
7.7.1 z-Transform solution of discrete-time state equations
Equation (7.73) can be obtained by z-transforming the discrete-time state equation
(7.57). The z-transform is given by
zXðzÞ 2 zxð0Þ 5 Ad XðzÞ 1 Bd UðzÞ
(7.75)
XðzÞ 5 ½zIn 2Ad 21 ½zxð0Þ 1 Bd UðzÞ
(7.76)
Hence, X(z) is given by
We therefore need to evaluate the inverse z-transform of the matrix [zIn 2 Ad]21z.
This can be accomplished by expanding the matrix in the form of the series
21
1
½zIn 2Ad 21 z 5 In 2 Ad
z
(7.77)
21
2 22
i 2i
.
.
.
.
.
.
5 In 1 A d z 1 A z 1 1 A z 1
d
d
The inverse z-transform of the series is
Zf½zIn 2 Ad 21 zg 5 fIn ; Ad ; A2d ; . . .; Aid ; . . .g
(7.78)
Hence, we have the z-transform pair
Z
½zIn 2Ad 21 z
!fAkd gN
k50
(7.79)
This result is analogous to the scalar transform pair
N
z
Z
! akd k50
z 2 ad
The inverse matrix in (7.79) can be evaluated using the Leverrier algorithm of
Section 7.4.1 to obtain the expression
½zIn 2Ad 21 z 5
P0 z 1 P1 z2 1 . . . 1 Pn21 zn
a0 1 a1 z 1 . . . 1 an21 zn21 1 zn
(7.80)
Then, after denominator factorization and partial fraction expansion, we obtain
½zIn 2Ad 21 z 5
n
X
i51
z
Zi
z 2 λi
(7.81)
where λi, i 5 1, 2, . . . , n are the eigenvalues of the discrete state matrix Ad.
Finally, we inverse z-transform to obtain the discrete-time state-transition matrix
Akd 5
n
X
Zi λki
(7.82)
i51
Writing (7.82) for k 5 1 and using (7.58), we have
Ad 5
n
X
i51
Zi λi 5
n
X
i51
Zi ðAÞeλi ðAÞT
(7.83)
7.7 Solution of discrete-time statespace equations
where the parentheses indicate terms pertaining to the continuous-time state
matrix A. Because the equality must hold for any sampling period T and any
matrix A, we have the two equalities
Zi 5 Zi ðAÞ
λi 5 eλi ðAÞT
(7.84)
In other words, the constituent matrices of the discrete-time state matrix are
simply those of the continuous-time state matrix A, and its eigenvalues are exponential functions of the continuous-time characteristic values times the sampling
period. This allows us to write the discrete-time state-transition matrix as
Akd 5
n
X
Zi eλi ðAÞkT
(7.85)
i51
For a state matrix with complex-conjugate eigenvalues λ1,2 5 σ 6 jωd, the state
transition matrix for the discrete-time system can be written as
Akd 5 eAkT 5 2 RefZgeσkT cosðωd kTÞ 2 2 ImfZgeσkT sinðωd kTÞ
(7.86)
To complete the solution of the discrete-time state equation, we examine the
zero-state response rewritten as
XZS ðzÞ 5 f½zIn 2 Ad 21 zgz21 Bd UðzÞ
(7.87)
The term in braces in (7.87) has a known inverse transform, and multiplication
by z21 is equivalent to delaying its inverse transform by one sampling period.
The remaining terms also have a known inverse transform. Using the convolution
theorem, the inverse of the product is the convolution summation
xZS ðkÞ 5
k21
X
Ak2i21
Bd uðiÞ
d
(7.88)
i50
This completes the solution using z-transformation.
Using (7.85), the zero-state response can be written as
xZS ðkÞ 5
"
k21 X
n
X
i50
#
Zj e
λj ðAÞhk2i21iT
(7.89)
Bd uðiÞ
j51
Then interchanging the order of summation gives
xZS ðkÞ 5
n
X
j51
"
Zj B d e
λj ðAÞhk21iT
k21
X
#
2λj ðAÞiT
e
uðiÞ
(7.90)
i50
This expression is useful in some special cases where the summation over i can
be obtained in closed form.
273
274
CHAPTER 7 StateSpace Representation
Occasionally, the initial conditions are given at nonzero time. The complete
solution (7.73) can be shifted to obtain the response
0
xðk0 Þ 1
xðkÞ 5 Ak2k
d
k21
X
Ak2i21
Bd uðiÞ
d
(7.91)
i5k0
where x(k0) is the state at time k0.
EXAMPLE 7.18
Consider the state equation
x_1
5
x_2
0
22
1
23
x1
0
1
u
x2
1
1. Solve the state equation for a unit step input and the initial condition vector x(0) 5 [1 0]T.
2. Use the solution to obtain the discrete-time state equations for a sampling period
of 0.1 s.
3. Solve the discrete-time state equations with the same initial conditions and input as
in part 1, and verify that the solution is the same as that shown in part 1 evaluated
at multiples of the sampling period T.
Solution
1. We begin by finding the state-transition matrix for the given system. The resolvent
matrix is
s13 1
1 0
3
1
21
s1
22 s
0 1
22 0
s 21
5
5 2
ΦðsÞ 5
2 s13
s 1 3s 1 2
ðs 1 1Þðs 1 2Þ
To inverse Laplace transform, we need the partial fraction expansions
1
1
21
5
1
ðs 1 1Þðs 1 2Þ ðs 1 1Þ ðs 1 2Þ
s
21
2
5
1
ðs 1 1Þðs 1 2Þ ðs 1 1Þ ðs 1 2Þ
Thus, we reduce the resolvent matrix to
3
1
3
1 0
1 0
1
2
2
2
22 0
22
0 1
0 1
1
ΦðsÞ 5
ðs 1 1Þ
ðs 1 2Þ
2
1
21 21
22 21
2
2
5
1
ðs 1 1Þ
ðs 1 2Þ
1
0
The matrices in the preceding expansion are both rank 1 because they are the constituent
matrices of the state matrix A. The expansion can be easily inverse Laplace transformed
to obtain the state-transition matrix
φðtÞ 5
2
22
1
e2t 1
21
21
2
21 22t
e
2
7.7 Solution of discrete-time statespace equations
The zero-input response of the system is
xZI ðtÞ 5
5
1
e22t
0
2
21
e2t 1
22 21
2
21
2
e22t
e2t 1
2
22
2
21
1
For a step input, the zero-state response is
xZS ðtÞ 5
5
2
1
e2t 1
22 21
1
e2t 1ðtÞ 1
21
21
0
1ðtÞ
e22t
1
2
21
2
21 22t
e
1ðtÞ
2
where * denotes convolution. Because convolution of an exponential and a step is a
relatively simple operation, we do not need Laplace transformation. We use the identity
e2αt 1ðtÞ 5
ðt
0
e2ατ dτ 5
1 2 e2αt
α
to obtain
xZS ðtÞ 5
1
ð1 2 e2t Þ 1
21
1=2
21 1 2 e22t
5
2
0
2
2
1
e2t 2
21
21 e22t
2
2
The total system response is the sum of the zero-input and zero-state responses
xðtÞ 5 xZI ðtÞ 1 xZS ðtÞ
#
"
#
"
# "
"
#
"
#
1
21 22t
1=2
21 e22t
2
e2t 2
e 1
2
e2t 1
5
2
21
2
0
2
22
#
"
#
"
# "
1
21 e22t
1=2
e2t 1
1
5
2
21
2
0
2. To obtain the discrete-time state equations, we use the state-transition matrix obtained
in step 1 with t replaced by the sampling period. For a sampling period of 0.1, we have
2
1
21 21 22 3 0:1
0:9909
0:0861
e20:1 1
e
5
Ad 5 Φð0:1Þ 5
22 21
2
2
20:1722 0:7326
The discrete-time input matrix Bd can be evaluated as shown earlier using (7.59).
One may also observe that the continuous-time system response to a step input of
the duration of one sampling period is the same as the response of a system due to a
piecewise constant input discussed in Section 7.6. If the input is of unit amplitude,
Bd can be obtained from the zero-state response of part 1 with t replaced by the
sampling period T 5 0.1. Bd is given by
9
8
< 2
1
21 21 1 2 e22 3 0:1 = 0
20:1
Bd 5
ð1 2 e Þ 1
; 1
: 22 21
2
2
2
20:2
1=2
1
0:0045
21 e
5
2
e20:1 2
5
2
0
21
0:0861
2
275
276
CHAPTER 7 StateSpace Representation
3. The solution of the discrete-time state equations involves the discrete-time state-transition
matrix
Akd 5 Φð0:1kÞ 5
1
e20:1k 1
21
2
22
21
2
21 20:2k
e
2
The zero-input response is the product
xZI ðkÞ 5 Φð0:1kÞ x ð0Þ
2
1
5
e20:1k 1
22 21
1
21 20:2k
5
e
0
2
21
2
2
e20:1k 1
22
21 20:2k
e
2
Comparing this result to the zero-input response of the continuous-time system reveals
that the two are identical at all sampling points k 5 0, 1, 2, . . ..
Next, we z-transform the discrete-time state-transition matrix to obtain
2
22
ΦðzÞ 5
z
1
1
21 z 2 e20:1
21
2
z
21
2 z 2 e20:2
Hence, the z-transform of the zero-state response for a unit step input is
XZS ðzÞ 5 ΦðzÞz21 Bd UðzÞ
9
8
= 0:0045 z21 z
< 2
21 21
1
z
z
1
5
20:1
20:2
; 0:0861 z 2 1
: 22 21 z 2 e
2
2 z2e
1
21
z
z
1 9:0635 3 1022
5 9:5163 3 1022
20:1
21 ðz 2 e Þðz 2 1Þ
2 ðz 2 e20:2 Þðz 2 1Þ
We now need the partial fraction expansions
z
ðz 2 e20:1 Þðz 2 1Þ
5
1
z
ð2 1Þz
z
ð2 1Þz
5
10:5083
1
1
1 2 e20:1 z 2 1 z 2 e20:1
z 2 1 z 2 0:9048
5
ð 21Þz
ð 21Þz
1
z
z
5
5:5167
1
1
1 2 e20:2 z 2 1 z 2 e20:2
z 2 1 z 2 0:8187
z
ðz 2 e20:2 Þðz 2 1Þ
Substituting in the zero-state response expression yields
Xzs ðzÞ 5
z
0:5
2
0 z21
z
z
21
1
2
0:5
2 z 2 e20:2
21 z 2 e20:1
Then inverse-transforming gives the response
xZS ðkÞ 5
0:5
2
0
1
21 20:2k
e20:1k 2 0:5
e
21
2
This result is identical to the zero-state response for the continuous system at time
t 5 0.1 k, k 5 0, 1, 2, . . .. The zero-state response can also be obtained using (7.90) and
the well-known summation
k21
X
i50
ai 5
1 2 ak
;
12a
jaj , 1
7.8 z-Transfer function from statespace equations
Substituting the summation in (7.90) with constant input and then simplifying the
result gives
2
3
n
2λj ðAÞkT
X
λj ðAÞik21hT 41 2 e
5
Zj Bd e
xZS ðkÞ 5
1 2 e2λj ðAÞT
j51
2
3
n
X
1 2 eλj ðAÞkT 5
4
Zj Bd
5
1 2 eλj ðAÞT
j51
Substituting numerical values in the preceding expression yields the zero-state response
obtained earlier.
7.8 z-Transfer function from statespace equations
The z-transfer function can be obtained from the discrete-time statespace representation by z-transforming the output equation (7.65) to obtain
YðzÞ 5 CXðzÞ 1 DUðzÞ
(7.92)
The transfer function is derived under zero initial conditions. We therefore
substitute the z-transform of the zero-state response (7.87) for X(z) to obtain
YðzÞ 5 Cf½zIn 2 Ad 21 gBd UðzÞ 1 DUðzÞ
(7.93)
Thus, the z-transfer function matrix is defined by
YðzÞ 5 GðzÞUðzÞ
where G(z) is the matrix
GðzÞ 5 Cf½zIn 2 Ad 21 gBd 1 D
Z
! GðkÞ 5
(7.94)
CAk21
Bd ;
d
D;
k$1
k50
(7.95)
and G(k) is the impulse response matrix. The transfer function matrix and the
impulse response matrix are z-transform pairs.
Substituting from (7.81) into (7.95) gives the alternative expression
8 n
n
<X
X
1
Z
CZi Bd λk21
; k$1
i
(7.96)
CZi Bd
1 D ! GðkÞ 5
GðzÞ 5
i51
:
z
2
λ
i
i51
D;
k50
Thus, the poles of the system are the eigenvalues of the discrete-time state matrix
Ad. From (7.84), these are exponential functions of the eigenvalues λi(A) of the
continuous-time state matrix A. For a stable matrix A, the eigenvalues λi(A) have
negative real parts and the eigenvalues λi have magnitude less than unity. This
implies that the discretization of Section 7.6 yields a stable discrete-time system
for a stable continuous-time system.
277
278
CHAPTER 7 StateSpace Representation
Another important consequence of (7.96) is that the product CZiB can vanish
and eliminate certain eigenvalues from the transfer function. This occurs if the
product CZi is zero, the product ZiB is zero, or both. If such cancellation occurs,
the system is said to have an output-decoupling zero at λi, an input-decoupling
zero at λi, or an input-output-decoupling zero at λi, respectively. The poles
of the reduced transfer function are then a subset of the eigenvalues of the state
matrix Ad. A statespace realization that leads to pole-zero cancellation is said to
be reducible or nonminimal. If no cancellation occurs, the realization is said
to be irreducible or minimal.
Clearly, in case of an output-decoupling zero at λi, the forced system response
does not include the mode λki . In case of an input-decoupling zero, the mode is
decoupled from or unaffected by the input. In case of an input-output-decoupling
zero, the mode is decoupled from both the input and the output. These properties are
related to the concepts of controllability and observability discussed in Chapter 8.
EXAMPLE 7.19
Obtain the z-transfer function for the position control system described in Example 7.18.
1. With x1 as output
2. With x1 1 x2 as output
Solution
From Example 7.18, we have
z
z
21 21
1
1
2
2 z 2 e20:2
21 z 2 e20:1
1
0:0045
1=2
21 e20:2
e20:1 2
2
5
Bd 5
21
0:0861
0
2
2
ΦðzÞ 5
2
22
1. The output matrix C and the direct transmission matrix D for output x1 are
C5 1
0
D50
Hence, the transfer function of the system is
GðzÞ 5 1
5
8
<
2
0
: 22
1
1
1
21 z 2 e20:1
21
2
9
= 0:0045 21
1
2 z 2 e20:2 ; 0:0861
9:5163 3 1022
9:0635 3 1022
2
z 2 e20:1
z 2 e20:2
2. With x1 1 x2 as output, C 5 [1, 1] and D 5 0. The transfer function is
GðzÞ 5 1
5
8
<
2
1
: 22
1
1
1
21 z 2 e20:1
0
9:0635 3 1022
1
z 2 e20:2
z 2 e20:1
21
2
9
=
1
0:0045
21
2 z 2 e20:2 ; 0:0861
7.9 Similarity transformation
The system has an output-decoupling zero at e20.1 because the product CZ1 is zero. The
response of the system to any input does not include the decoupling term. For example,
the step response is the inverse of the transform
9:0635 3 1022 z
ðz 2 e20:2 Þðz 2 1Þ
2
3
2
3
9:0635 3 1022 4 z
ð2 1Þz 5
z
ð2 1Þz 5
4
5
1
1
5 0:5
1 2 e20:2
z 2 1 z 2 e20:2
z 2 1 z 2 0:8187
Y ðzÞ 5
That is, the step response is
yðkÞ 5 0:5½1 2 e20:2k 7.8.1 z-Transfer function in MATLAB
The expressions for obtaining z-domain and s-domain transfer functions differ
only in that z in the former is replaced by s in the latter. The same MATLAB
command is used to obtain s-domain and z-domain transfer functions. The transfer
function for the matrices (Ad, Bd, C, D) is obtained with the commands
.. p 5 ssðAd;Bd;C;D;TÞ
.. gd5 tfðpÞ
where T is the sampling period. The poles and zeros of a transfer function are
obtained with the command
.. zpkðgdÞ
For the system described in Example 7.19 (2), we obtain
Zero/pole/gain:
0.090635 (z 2 0.9048)
––––––––––––––
(z 2 0.9048) (z 2 0.8187)
Sampling time: 0.1
The command reveals that the system has a zero at 0.9048 and poles at (0.9048,
0.8187) with a gain of 0.09035. With pole-zero cancellation, the transfer function
is the same as that shown in Example 7.19 (2).
7.9 Similarity transformation
Any given linear system has an infinite number of valid statespace representations.
Each representation corresponds to a different set of basis vectors in statespace.
Some representations are preferred because they reveal certain system properties,
whereas others may be convenient for specific design tasks. This section considers
transformation from one representation to another.
279
280
CHAPTER 7 StateSpace Representation
Given a state vector x(k) with statespace representation of the form (7.57),
we define a new state vector z(k)
xðkÞ 5 Tr zðkÞ3zðkÞ 5 Tr21 xðkÞ
(7.97)
where the transformation matrix Tr is assumed invertible. Substituting for x(k)
from (7.97) in the state equation (7.57) and the output equation (7.65) gives
Tr zðk 1 1Þ5 Ad Tr zðkÞ 1 Bd uðkÞ
(7.98)
yðkÞ 5 CTr zðkÞ 1 DuðkÞ
Premultiplying the state equation by Tr21 gives
zðk 1 1Þ 5 Tr21 Ad Tr zðkÞ 1 Tr21 Bd uðkÞ
(7.99)
Hence, we have the statespace quadruple for the state vector z(k)
ðA; B; C; DÞ 5 ðTr21 Ad Tr ; Tr21 Bd ; CTr ; DÞ
(7.100)
Clearly, the quadruple for the state vector x(k) can be obtained from the quadruple
of z(k) using the inverse of the transformation Tr (i.e., the matrix Tr21 ).
For the continuous-time system of (7.3), a state vector z(t) can be defined as
xðtÞ 5 Tr zðtÞ3zðtÞ 5 Tr21 xðtÞ
(7.101)
and substitution in (7.3) yields (7.100). Thus, a discussion of similarity transformation is identical for continuous-time and discrete-time systems. We therefore
drop the subscript d in the sequel.
EXAMPLE 7.20
Consider the point mass m driven by a force f of Example 7.1, and determine the two
statespace equations when m 5 1 and
1. The displacement and the velocity are the state variables
2. The displacement and the sum of displacement and velocity are the state variables
Solution
1. From the first principle, as shown in Example 7.1, we have that the equation of motion is
m€y 5 f
and in case 1, by considering x1 as the displacement and x2 as the velocity, we can
write
x_1 ðtÞ 5 x2 ðtÞ
x_2 ðtÞ 5 uðtÞ
yðtÞ 5 x1 ðtÞ
Thus, we obtain the realization
A5
1
0
0
0
B5
0
1
C5 1
0
D50
7.9 Similarity transformation
2. We express the new state variables, z1 and z2, in terms of the state variable of part 1 as
z1 ðtÞ 5 x1 ðtÞ
z2 ðtÞ 5 x1 ðtÞ 1 x2 ðtÞ
This yields the inverse of the transformation matrix
1 0
Tr21 5
1 1
Using (7.100) gives the realization
A 5 Tr21 Ad Tr 5
C5 1
21 1
21 1
0
B 5 Tr21 Bd 5
0
1
D50
To obtain the transformation to diagonal form, we recall the expression
A 5 VΛV 21 3Λ 5 V 21 AV
(7.102)
where V is the modal matrix of eigenvectors of A and Λ 5 diag{λ1, λ2, . . . , λn} is
the matrix of eigenvalues of A. Thus, for A 5 Λ in (7.100), we use the modal matrix
of A as the transformation matrix. The form thus obtained is not necessarily the
same as the diagonal form obtained in Section 8.5.3 from the transfer function using
partial fraction expansion. Even though all diagonal forms share the same state
matrix, their input and output matrices may be different.
EXAMPLE 7.21
Obtain the diagonal form for the statespace equations
2
3 2
32
3 2 3
0
1
0
x1 ðkÞ
0
x1 ðk 1 1Þ
4 x2 ðk 1 1Þ 5 5 4 0
0
1 54 x2 ðkÞ 5 1 4 0 5uðkÞ
x3 ðk 1 1Þ
x3 ðkÞ
0 20:04 20:5
1
2
2
3
x1 ðkÞ
yðkÞ 5 4 1 0 0 4 x2 ðkÞ 5
x3 ðkÞ
Solution
The eig command of MATLAB yields the eigenvalues and the modal matrix whose columns
have unity norm
2
3
1
20:995
0:9184
4
Λ 5 diagf0; 2 0:1; 2 0:4g V 5 0
0:0995
20:3674 15
0 20:00995
0:1469
The state matrix is in companion form and the modal matrix is also known to be the
Van der Monde matrix (column norms need not be unity):
2
1
V 5 4 λ1
λ21
1
λ2
λ22
3 2
1
1
λ3 5 5 4 0
λ23
0
1
20:1
0:01
3
1
20:4 5
0:16
281
282
CHAPTER 7 StateSpace Representation
The MATLAB command for similarity transformation is ss2ss. It requires the inverse Tr21 of
the similarity transformation matrix Tr. Two MATLAB commands transform to diagonal
form. The first requires the modal matrix of eigenvectors V and the system to be
transformed
.. s diag 5 ss2ssðsystem; invðvÞÞ
The second uses similarity transformation but does not require the transformation matrix
.. s diag 5 cannonðsystem; 'modal'Þ
The two commands yield
At 5 diagf0; 2 0:1; 2 0:4g;
Bt 5 ½25 33:5012 9:0738T
Ct 5 ½1:0000
Dt 5 0
2 0:9950 0:9184
For complex conjugate eigenvalues, the command canon yields a real realization but its
state matrix is not in diagonal form, whereas ss2ss will yield a diagonal but complex
matrix.
7.9.1 Invariance of transfer functions and characteristic equations
Similar systems can be viewed as different representations of the same systems.
This is justified by the following theorem.
THEOREM 7.1
Similar systems have identical transfer functions and characteristic polynomials.
PROOF
Consider the characteristic polynomials of similar realizations (A, B, C, D) and (A1, B1, C1, D):
detðzIn 2 A1 Þ 5 detðzIn 2 Tr21 ATr Þ
5 det½Tr21 ðzIn 2 AÞTr 5 det½Tr21 det ðzIn 2 AÞ det½Tr 5 detðzIn 2 AÞ
where we used the identity det Tr21 3 det½Tr 5 1.
The transfer function matrix is
G1 ðsÞ 5 C1 ½zIn 2 A1 B1 1 D1
5 CTr ½zIn 2 Tr21 ATr Tr21 B 1 D
5 C½Tr ðzIn 2 Tr21 ATr ÞTr21 B 1 D
5 C½zIn 2 AB 1 D 5 GðsÞ
where we used the identity (A B C)21 5 C21 B21 A21.’
Clearly, not all systems with the same transfer function are similar, due to the
possibility of pole-zero cancellation. Systems that give rise to the same transfer
function are said to be equivalent.
Problems
EXAMPLE 7.22
Show that the following system is equivalent to the system shown in Example 7.19 (2).
xðk 1 1Þ 5 0:8187xðkÞ 1 9:0635 3 1022 uðkÞ
yðkÞ 5 xðkÞ
Solution
The transfer function of the system is
GðzÞ 5
9:0635 3 1022
z 2 0:8187
which is identical to the reduced transfer function of Example 7.19 (2).
Resources
D’Azzo, J.J., Houpis, C.H., 1988. Linear Control System Analysis and Design. McGraw-Hill,
New York.
Belanger, P.R., 1995. Control Engineering: A Modern Approach. Saunders, Fort Worth,
TX.
Brogan, W.L., 1985. Modern Control Theory. Prentice Hall, Englewood Cliffs, NJ.
Chen, C.T., 1984. Linear System Theory and Design. HRW, New York.
Friedland, B., 1986. Control System Design: an Introduction to StateSpace Methods.
McGraw-Hill, New York.
Gupta, S.C., Hasdorff, L., 1970. Fundamentals of Automatic Control. Wiley, New York.
Hou, S.-H., 1998. A simple proof of the Leverrier-Faddeev characteristic polynomial algorithm.
SIAM Review 40 (3), 706709.
Kailath, T., 1980. Linear Systems. Prentice Hall, Englewood Cliffs, NJ.
Moler, C.B., Van Loan, C.F., 1978. Nineteen dubious ways to calculate the exponential
of a matrix. SIAM Review 20, 801836.
Sinha, N.K., 1988. Control Systems. HRW, New York.
PROBLEMS
7.1
Classify the statespace equations regarding linearity and time variance:
"
(a)
x_1
x_2
#
"
5
sinðtÞ
y5½1
0
1
1
22
" #
x1
x2
#"
x1
x2
#
1
" #
1
u
2
283
284
CHAPTER 7 StateSpace Representation
2
3 2
1
x_1
(b) 4 x_2 5 5 4 0
x_3
1
y5½1
32 3 2 3
1 0
1
x1
2 0 54 x2 5 1 4 2 5u
x3
5 7
0
2 3
x1
1 2 4 x2 5
x3
(c) x_ 5 22x2 1 7x 1 xu
y 5 3x
(d) x_ 5 27x 1 u
y 5 3x2
7.2
The equations of motion of a 2-D.O.F. manipulator are
T
_
€
M θ 1 DðθÞ 1 gðθÞ 5
f
m11 m12
0
g1 ðθÞ
_ 5
M5
dðθÞ
gðθÞ
5
m12 m22
D2 θ_ 2
g2 ðθÞ
ma11 ma12
Ma 5 M 21 5
ma12 ma22
where θ 5 [θ1, θ2]T is a vector of joint angles. The entries of the positive
definite inertia matrix M depend on the robot coordinates θ. D2 is a damping
constant. The terms gi, i 5 1, 2, are gravity-related terms that also depend
on the coordinates. The right-hand side is a vector of generalized forces.
(a) Obtain a statespace representation for the manipulator, and then
linearize it in the vicinity of a general operating point (x0, u0).
(b) Obtain the linearized model in the vicinity of zero coordinates, velocities,
and inputs.
(c) Show that, if the entries of the state matrix are polynomials, the answer
to (b) can be obtained from (a) by letting all nonlinear terms go to zero.
7.3
Obtain the matrix exponentials for the state matrices using four different
approaches:
(a) A 5 diagf 23; 25; 27g
2
3
0
0
1
(b) A 5 4 0
21 0 5
26
0
0
Problems
2
0
(c) A 5 4 0
0
1
0
26
3
1
1 5
25
(d) A is a block diagonal matrix with the matrices of (b) and (c) on its diagonal.
7.4
Obtain the zero-input responses of the systems of Problem 7.3 due to the
initial condition vectors:
(a), (b), (c) [1, 1, 0]T and [1, 0, 0]T
(d) [1, 1, 0, 1, 0, 0]T
7.5
Determine the discrete-time state equations for the systems of Problem 7.3
(a), (b), and (c), with b 5 [0, 0, 1]T in terms of the sampling period T.
7.6
Prove that the (right) eigenvectors of the matrix AT are the left eigenvectors
of A and that AT’s eigenvalues are the eigenvalues of A using
A 5 VΛV 21 5 VΛW
7.7
Prove the properties of the constituent matrices given in Section 7.4.3
using (7.44).
7.8
(a) Derive the expressions for the terms of the adjoint matrix used in the
Leverrier algorithm. Hint: Multiply both sides of (7.28) by the matrix
[sI 2 A] and equate coefficients.
(b) Derive the expressions for the coefficients of the characteristic equations
used in the Leverrier algorithm. Hint: Laplace transform the derivative
expression for the matrix exponential to obtain
s L feAt g 2 I 5 L feAt gA
Take the trace, and then use the identity
tr sIn 2 A21
7.9
5
a1 1 2a2 s 1 . . . 1ðn 2 1Þan21 sn22 1 nsn21
a0 1 a1 s 1 . . . 1 an21 sn21 1 sn
The biological component of a fishery system is assumed to be governed
by the population dynamics equation
dxðtÞ
5 rxðtÞð1 2 xðtÞ=KÞ 2 hðtÞ
dt
where r is the intrinsic growth rate per unit time, K is the environment carrying capacity, x(t) is the stock biomass, and h(t) is the harvest rate in weight.2
(a) Determine the harvest rate for a sustainable fish population x0 , K.
(b) Linearize the system in the vicinity of the fish population x0.
2
Clark, C.W., 1990. Mathematical Bioeconomics: The Optimal Management of Renewable
Resources. Wiley, New York.
285
286
CHAPTER 7 StateSpace Representation
(c) Obtain a discrete-time model for the linearized model with a fixed
average yearly harvest rate h(k) in the kth year.
(d) Obtain a condition for the stability of the fish population from your
discrete-time model, and comment on the significance of the condition.
7.10
The following differential equations represent a simplified model of an
overhead crane:3
ðmL 1 mC Þ€x1 ðtÞ 1 mL lð€x3 ðtÞ cos x3 ðtÞ 2 x_23 ðtÞ sin x3 ðtÞÞ 5 u
mL x€ 1 ðtÞ 1 cosx3 ðtÞ 1 mL l€x3 ðtÞ 5 2 mL g sin x3 ðtÞ
where mC is the mass of the trolley, mL is the mass of the hook/load, l is the
rope length, g is the gravity acceleration, u is the force applied to the trolley,
x1 is the position of the trolley, and x3 is the rope angle. Consider the position
of the load y 5 x1 1 l sin x3 as the output.
(a) Determine a linearized statespace model of the system about the
equilibrium point x 5 0 with state variables x1, x3, the first derivative
of x1, and the first derivative of x3.
(b) Determine a second statespace model when the sum of the trolley
position and of the rope angle is substituted for the rope angle as a
third state variable.
7.11
Obtain the diagonal form for the discretized armature-controlled DC motor
system of Example 7.15 with the motor angular position as output.
7.12
A system whose state and output responses are always nonnegative for
any nonnegative initial conditions and any nonnegative input is called a positive system.4 Positive systems arise in many applications where the system
variable can never be negative, including chemical processes, biological systems, and economics, among others. Show that the single-input-single-output
discrete-time system (A, b, cT) is positive if and only if all the entries of the
state, input, and output matrix are positive.
7.13
To monitor river pollution, we need to model the concentration of biodegradable matter contained in the water in terms of biochemical oxygen demand
for its degradation. We also need to model the dissolved oxygen deficit
defined as the difference between the highest concentration of dissolved oxygen and the actual concentration in mg/l. If the two variables of interest are the
state variables x1 and x2, respectively, then an appropriate model is given by
0
2k1
_x 5
x
2k2
k1
3
Piazzi A., Visioli, A., 2002. Optimal dynamic-inversion-based control of an overhead crane. IEE
Proceedings: Control Theory and Applications 149 (5), 405411.
4
Farina, L., Rinaldi, S., 2000. Positive Linear Systems: Theory & Applications. Wiley-Interscience,
New York.
Problems
where k1 is a biodegradation constant and k2 is a reaeration constant, and
both are positive. Assume that the two positive constants are unequal.
Obtain a discrete-time model for the system with sampling period T, and
show that the system is positive.
7.14
Autonomous underwater vehicles (AUVs) are robotic submarines that
can be used for a variety of studies of the underwater environment.
The vertical and horizontal dynamics of the vehicle must be controlled
to remotely operate the AUV. The INFANTE (Figure P7.14) is a research
AUV operated by the Instituto Superior Tecnico of Lisbon, Portugal.5
The variables of interest in horizontal motion are the sway speed and
the yaw angle. A linearized model of the horizontal plane motion of the
vehicle is given by
2 3 2
32 3 2
3
20:14 20:69 0:0
0:056
x1
x_1
4 x_2 5 5 4 20:19 20:048 0:0 54 x2 5 1 4 20:23 5u
x3
x_3
0:0
1:0
0:0
0:0
2 3
x
1 0 0 4 15
y1
5
x2
y2
0 1 0
x3
where x1 is the sway speed, x2 is the yaw angle, x3 is the yaw rate, and
u is the rudder deflection. Obtain the discrete statespace model for the
system with a sampling period of 50 ms.
FIGURE P7.14
The INFANTE AUV. (From Silvestrea and Pascoa, 2004; used with permission.)
5
Silvestrea, C., Pascoa, A., 2004. Control of the INFANTE AUV using gain scheduled static output
feedback. Control Engineering Practice 12, 15011509.
287
288
CHAPTER 7 StateSpace Representation
7.15
A simplified linear state-space model of the ingestion of a drug in the
bloodstream is given by6
0
2K1
x1
1
x_1
5
1
u
2K2 x2
K1
x_2
0
where
x1 5 the mass of the absorbed drug in mg
x2 5 the mass of the drug in the bloodstream in mg
u 5 the rate at which the drug is ingested in mg/min
(a) Find the state-transition matrix of the system.
(b) Obtain a discrete-time state-space model for the system with a sampling
period T in terms of the model parameters.
7.16
A typical assumption in most mathematical models is that the system
differential equations or transfer functions have real coefficients. In a few
applications, this assumption is not valid. In models of rotating machines,
the evolution of the vectors governing the system with time depends on
their space orientation relative to fixed inertial axes. For an induction
motor, assuming symmetry, two stator fixed axes are used as the reference
frame: the direct axis (d) in the horizontal direction and the quadrature axis (q)
in the vertical direction. The terms in the quadrature direction are identified
with a (j) coefficient that is absent from the direct axis terms. The two axes are
shown in Figure P7.15.
q
stator
d
rotor
FIGURE P7.15
Stator frame for the induction motor.
6
McClamroch, N.H., 1980. State Models of Dynamic Systems: A Case Study Approach. SpringerVerlag.
Computer Exercises
We write the equations for the electrical subsystem of the motor in terms
of the stator and rotor currents and voltages. Each current is decomposed
into a direct axis component and quadrature component, with the latter
identified with the term (j). The s-domain equations of the motor are
obtained from its equivalent circuit using Kirchhoff’s laws. The equations
relative to the stator axes and including complex terms are
Rs 1 sLs
vs
5
vr
ðs 2 jωr ÞLm
sLm
Rr 1ðs 2 jωr ÞLr
is
ir
where
Rs (Rr) 5 stator (rotor) resistance
Ls, Lr, Lm 5 stator, rotor, mutual inductance, respectively
ωr 5 rotor angular velocity
(a) Write the state equations for the induction motor.
(b) Without obtaining the eigenvalues of the state matrix, show that the
two eigenvalues are not complex conjugate.
7.17
In many practical applications, the output sampling in a digital control
system is not exactly synchronized with the input transition. Show that
the output equation corresponding to output sampling at tk 5 kT 1 ∆k,
k 5 0,1,2, . . . is
yðkÞ 5 CðkÞxðkÞ 1 DðkÞuðkÞ
Ð∆
where CðkÞ 5 CeA∆k ; DðkÞ 5 CBd ð∆k Þ 1 D; Bd ð∆k Þ 5 0 k eAt Bdτ:
• If the direct transmission matrix D is zero, does the input directly influence
the sampled output?
• When is the resulting input-output model time-invariant?
COMPUTER EXERCISES
7.18
Write a computer program to simulate the systems of Problem 7.1 for various initial conditions with zero input, and discuss your results referring to
the solutions of Example 7.4. Obtain plots of the phase trajectories for any
second-order system.
7.19
Write a program to obtain the state-transition matrix using the Leverrier
algorithm.
7.20
Simulate the systems of Problem 7.3 (ac) with the initial conditions of 7.4,
and obtain statetrajectory plots with one state variable fixed for each system.
289
290
CHAPTER 7 StateSpace Representation
7.21
Repeat Problem 7.5 using a computer-aided design (CAD) package for two
acceptable choices of the sampling period and compare the resulting systems.
7.22
Simulate the river pollution system of Problem 7.13 for the normalized
parameter values of k1 5 1, k2 5 2, with a sampling period T 5 0.01 s for
the initial conditions xT (0) 5 [1, 0], [0, 1], [1, 1], and plot all the results
together.
7.23
Repeat Problem 7.14 using a CAD package.
7.24
By evaluating the derivatives of the matrix exponential at t 5 0, show that
for any matrix A with distinct eigenvalues the constituent matrices can be
obtained by solving the equation
2
In
6 λ1 In
6
4 ^
λn21
1 In
In
λ2 In
^
λn21
2 In
?
?
&
?
32 3 2
3
In
Z1
In
6 7 6
7
λn In 7
76 Z2 7 5 6 A 7
^ 54 ^ 5 4 ^ 5
λn21
Zn
An21
n In
Write a MATLAB function that evaluates the constituent matrices and
save them in a cell array.
7.25
The Cayley-Hamilton theorem states that, if f(λ) 5 λn 1 an21λn21 1 ?
1 a1λ 1 a0 is the characteristic polynomial of a matrix A, then f(A) 5
An 1 an21An21 1 ? 1 a1A 1 a0In 5 0.
This allows us to express the nth power as
An 5 2½an21 An21 1 ? 1 a1 A 1 a0 In (a) Verify the validity of the theorem for the matrix
2
3
A548
0
22
23
4
3
0
24 5
29
using the MATLAB commands poly and polyvalm. Note that the
answer you get will not be exact because of numerical errors.
(b) Use the Cayley-Hamilton theorem to show that the matrix exponential
of a matrix A can be written as
eAt 5
n21
X
αðtÞAi
i50
(c) Show that the initial value of the vector of time functions
Computer Exercises
αðtÞ 5 α0 ðtÞ
. . . αn21 ðtÞ T
is given by
T
αð0Þ 5 1 01 3 ðn21Þ
(d) By differentiating the expression of (b), use the Cayley-Hamilton
theorem to show that the vector of time functions satisfies
0T T
_ 5
2
a
αðtÞ
αðtÞ
In21 (e) The transpose of the matrix of (d) is in a companion form whose
eigenvector decomposition can be written by inspection. Use this fact
to obtain the equation
2
3
2 λt3
e 1
1 λ1 ? λn21
1
4^ ^ &
^ 5αðtÞ 5 4 ^ 5
e λn t
1 λn ? λn21
n
(f) Write a MATLAB function to calculate the matrix exponential using
the Cayley-Hamilton theorem and the results of (be).
291
CHAPTER
Properties of StateSpace
Models
8
OBJECTIVES
After completing this chapter, the reader will be able to do the following:
1. Determine the equilibrium point of a discrete-time system.
2. Determine the asymptotic stability of a discrete-time system.
3. Determine the input-output stability of a discrete-time system.
4. Determine the poles and zeros of multivariable systems.
5. Determine controllability and stabilizability.
6. Determine observability and detectability.
7. Obtain canonical statespace representations from an input-output representation
of a system.
In Chapter 7, we described statespace models of linear discrete time and how
these models can be obtained from transfer functions or input-output differential
equations. We also obtained the solutions to continuous time and discrete-time
state equations. Now, we examine some properties of these models that play an
important role in system analysis and design.
We examine controllability, which determines the effectiveness of state feedback control; observability, which determines the possibility of state estimation
from the output measurements; and stability. These three properties are independent, so a system can be unstable but controllable, uncontrollable but stable, and
so on. However, systems whose uncontrollable dynamics are stable are stabilizable, and systems whose unobservable dynamics are stable are called detectable.
Stabilizability and detectability are more likely to be encountered in practice
than controllability and observability. Finally, we show how statespace representations of a system in several canonical forms can be obtained from its
input-output representation.
To simplify our notation, the subscript d used in Chapter 7 with discrete-time
state and input matrices is dropped if the discussion is restricted to discrete-time
systems. In sections involving both continuous-time and discrete-time systems,
the subscript d is retained. We begin with a discussion of stability.
293
294
CHAPTER 8 Properties of StateSpace Models
8.1 Stability of statespace realizations
The concepts of asymptotic stability
(BIBO) stability of transfer functions are
more complete coverage using statespace
stability of statespace realizations. Then
input-output responses.
and bounded-input-bounded-output
discussed in Chapter 4. Here, we give
models. We first discuss the asymptotic
we discuss the BIBO stability of their
8.1.1 Asymptotic stability
The natural response of a linear system because of its initial conditions may:
1. Converge to the origin asymptotically
2. Remain in a bounded region in the vicinity of the origin
3. Grow unbounded
In the first case, the system is said to be asymptotically stable; in the second, the
system is marginally stable; and in the third, it is unstable. Clearly, physical
variables are never actually unbounded even though linear models suggest this.
Once a physical variable leaves a bounded range of values, the linear model
ceases to be valid and the system must be described by a nonlinear model.
Critical to the understanding of stability of both linear and nonlinear systems is
the concept of an equilibrium state.
DEFINITION 8.1: EQUILIBRIUM
An equilibrium point or state is an initial state from which the system never departs unless
perturbed.
For the state equation
x ðk 1 1Þ 5 f½x ðkÞ
(8.1)
all equilibrium states xe satisfy the condition
x ðk 1 1Þ 5 f½xðkÞ
5 f½xe 5 xe
(8.2)
For linear state equations, equilibrium points satisfy the condition
xðk 1 1Þ 5 AxðkÞ
5 Axe 5 xe 3½A 2 In xe 5 0
(8.3)
For an invertible matrix A 2 In, (8.3) has the unique solution 0, and the linear
system has one equilibrium state at the origin. Later we show that invertibility of
A 2 In is a necessary condition for asymptotic stability.
Unlike linear systems, nonlinear models may have several equilibrium states.
Leaving one equilibrium state, around which the linear model is valid, may drive
8.1 Stability of statespace realizations
the system to another equilibrium state where another linear model is required.
Nonlinear systems may also exhibit other more complex phenomena that are not
discussed in this chapter but are discussed in Chapter 11.
The equilibrium of a system with constant input u can be derived from
Definition 8.1 by first substituting the constant input value in the state equation to
obtain the form of equation (8.2).
EXAMPLE 8.1
Find the equilibrium points of the following systems:
1. x(k 1 1) 5 x(k) [x(k) 2 0.5]
2. x(k 1 1) 5 2x(k)
0:1
x1 ðk 1 1Þ
5
3.
1
x2 ðk 1 1Þ
0
0:9
x1 ðkÞ
x2 ðkÞ
Solution
1. At equilibrium, we have
xe 5 xe ½xe 2 0:5
We rewrite the equilibrium condition as
xe ½xe 2 1:5 5 0
Hence, the system has the two equilibrium states
xe 5 0 and
xe 5 1:5
2. The equilibrium condition is
xe 5 2xe
The system has one equilibrium point at xe 5 0.
3. The equilibrium condition is
0:1
x1e ðkÞ
5
1
x2e ðkÞ
0
0:9
x1e ðkÞ
0:1 2 1
0
x1e ðkÞ
0
3
5
1
0:9 2 1 x2e ðkÞ
0
x2e ðkÞ
The system has a unique equilibrium state at xe 5 [x1e, x2e]T 5 [0, 0]T.
Although the systems in Example 8.1 all have equilibrium points, convergence
to an equilibrium point is not guaranteed. This additional property defines asymptotic stability for linear systems assuming that the necessary condition of a unique
equilibrium at the origin is satisfied.
DEFINITION 8.2: ASYMPTOTIC STABILITY
A linear system is said to be asymptotically stable if all its trajectories converge to the
origin—that is, for any initial state x(k0), x(k)-0 as k-N.
295
296
CHAPTER 8 Properties of StateSpace Models
THEOREM 8.1
A discrete-time linear system is asymptotically (Schur) stable if and only if all the eigenvalues
of its state matrix are inside the unit circle.
PROOF
To prove the theorem, we examine the zero-input response in (7.73) with the state-transition
matrix given by (7.82). Substitution yields
xZI ðkÞ 5 Ak2k0 xðk0 Þ
n
X
0
Zi xðk0 Þλk2k
5
i
(8.4)
i51
Sufficiency
The response decays to zero if the eigenvalues are all inside the unit circle. Hence, the
condition is sufficient.
Necessity
To prove necessity, we assume that the system is stable but that one of its eigenvalues λj is
outside the unit circle. Let the initial condition vector x(k0) 5 vj, the jth eigenvector of A.
Then (see Section 7.4.3) the system response is
xZI ðkÞ 5
n
X
0
Zi xðk0 Þλk2k
i
i51
5
n
X
0
0
vi wTi vj λk2k
5 vj λk2k
i
j
i51
which is clearly unbounded as k-N. Hence, the condition is also necessary by contradiction.
REMARKS
For some nonzero initial states, the product Zj x(k0) is zero because the matrix Zj is rank 1
(x(k0) 5 vi, i 6¼ j). So the response to those initial states may converge to zero even if the
corresponding λj has magnitude greater than unity. However, the solution grows unbounded
for an initial state in the one direction for which the product is nonzero. Therefore, not all
trajectories converge to the origin for jλjj . 1. Hence, the condition of Theorem 8.1 is
necessary.
We now discuss the necessity of an invertible matrix A 2 In for asymptotic
stability. The matrix can be decomposed as
A 2 In 5 V½Λ 2 In V 21 5 V diagfλi 2 1gV 21
An asymptotically stable matrix A has no unity eigenvalues, and A 2 In is invertible.
EXAMPLE 8.2
Determine the stability of the systems of Example 8.1 (2) and 8.1 (3) using basic principles,
and verify your results using Theorem 8.1.
8.1 Stability of statespace realizations
Solution
8.1 (2): Consider any initial state x(0). Then the response of the system is
xð1Þ 5 2xð0Þ
xð2Þ 5 2xð1Þ 5 2 3 2xð0Þ 5 4xð0Þ
xð3Þ 5 2xð2Þ 5 2 3 4xð0Þ 5 8xð0Þ
^
xðkÞ 5 2k xð0Þ
Clearly, the response is unbounded as k-N. Thus, the system is unstable. The system
has one eigenvalue at 2 .1, which violates the stability condition of Theorem 8.1.
8.1 (3): The state equation is
0:1 0
x1 ðkÞ
x1 ðk 1 1Þ
5
1 0:9 x2 ðkÞ
x2 ðk 1 1Þ
Using Sylvester’s expansion (7.34), we obtain the constituent matrices
0:1 0
1
0
0
0
5
0:1 1
0:9
1 0:9
21=0:8 0
1=0:8 1
From (7.73), the response of the system due to an arbitrary initial state x(0) is
x1 ðkÞ
1
0
x1 ð0Þ
0
0
5
ð0:1Þk 1
ð0:9Þk
x2 ðkÞ
x2 ð0Þ
21=0:8 0
1=0:8 1
This decays to zero as k-N. Hence, the system is asymptotically stable. Both system
eigenvalues (0.1 and 0.9) are inside the unit circle, and the system satisfies the conditions
of Theorem 8.1.
Note that the response due to any initial state with x1(0) 5 0 does not include the first
eigenvalue. Had this eigenvalue been unstable, the response due to this special class of initial
conditions would remain bounded. However, the response due to other initial conditions would
be unbounded, and the system would be unstable.
In Chapter 4, asymptotic stability was studied in terms of the poles of the system with the additional condition of no pole-zero cancellation. The condition for
asymptotic stability was identical to the condition imposed on the eigenvalues of
a stable state matrix in Theorem 8.1. This is because the eigenvalues of the state
matrix A and the poles of the transfer function are identical in the absence of
pole-zero cancellation. Next, we examine a stability definition based on the inputoutput response.
8.1.2 BIBO stability
For an input-output system description, the system output must remain bounded for
any bounded-input function. To test the boundedness of an n-dimensional output
vector, a measure of the vector length or size known as the norm of the vector
must be used (see Appendix III). A vector x is bounded if it satisfies the condition
:x: , bx , N
for some finite constant bx where :x: denotes any vector norm.
(8.5)
297
298
CHAPTER 8 Properties of StateSpace Models
For real or complex n 3 1 vectors, all norms are equivalent in the sense that
if a vector has a bounded norm :x:a, then any other norm :x:b, is also bounded.
This is true because for any two norms, :x:a and :x:b there exist finite positive
constants k1 and k2 such that
k1 :x:a # :x:b # :x:a k2
(8.6)
Because a change of norm merely results in a scaling of the constant bx, boundedness is independent of which norm is actually used in (8.5). Stability of the
input-output behavior of a system is defined as follows.
DEFINITION 8.3: BOUNDED-INPUT-BOUNDED-OUTPUT STABILITY
A system is BIBO stable if its output is bounded for any bounded input. That is,
:uðkÞ: , bu , N.:yðkÞ: , by , N
(8.7)
To obtain a necessary and sufficient condition for BIBO stability, we need the
concept of the norm of a matrix (see Appendix III). Matrix norms can be defined
based on a set of axioms or properties. Alternatively, we define the norm of a
matrix in terms of the vector norm as
:A: 5 max
x
:Ax:
:x:
(8.8)
5 max :Ax:
:x:51
Multiplying by the norm of x, we obtain the inequality
:Ax: # :A:U:x:
(8.9)
which applies to any matrix norm induced from a vector norm. Next, we give a
condition for BIBO stability in terms of the norm of the impulse response
matrix.
THEOREM 8.2
A system is BIBO stable if and only if the norm of its impulse response matrix is absolutely
summable. That is,
N
X
k50
:GðkÞ: , N
(8.10)
8.1 Stability of statespace realizations
PROOF
Sufficiency
The output of a system with impulse response matrix G(k) and input u(k) is
i
X
yðiÞ 5
GðkÞuði 2 kÞ
k50
The norm of y(i) satisfies
X
i
:yðiÞ: 5 GðkÞuði 2 kÞ
k50
#
i
i
X
X
:GðkÞuði 2 kÞ: #
:GðkÞ: :uði 2 kÞ:
k50
k50
For a bounded input u(k)
:uðkÞ: , bu , N
Hence, the output is bounded by
:yðiÞ: # bu
i
X
:GðkÞ: # bu
k50
N
X
:GðkÞ:
k50
which is finite if condition (8.10) is satisfied.
Necessity
The proof is by contradiction. We assume that the system is BIBO stable but that (8.10) is
violated. We write the output y(k) in the form
yðkÞ 5 GðkÞu
3
uðkÞ
6 7
7
6
6 uðk 2 1Þ 7
7
6
7
GðkÞ 6
7
6
^
7
6
7
6
4 5
uð0Þ
5 Gð0Þ Gð1Þ
...
2
The norm of the output vector can be written as
:yðkÞ: 5 maxjys ðkÞj
s
5 maxjgTs ðkÞuj
s
gTs ðkÞ
where
is the s row of the matrix GðkÞ. Select the vectors uðiÞ to have the rth entry with
unity magnitude and with sign opposite to that of gsr(i), where gsr(i) denotes the srth entry
of the impulse response matrix G(i). Using the definition of a matrix norm as the maximum
row sum, this gives the output norm
:yðkÞ: 5 max
s
k X
m
X
i50 r50
jgsr ðiÞj
299
300
CHAPTER 8 Properties of StateSpace Models
For BIBO stability, this sum remains finite. But for the violation of (8.10),
k
X
:GðkÞ: 5
i50
k X
l X
m
X
jgsr ðiÞj-N
as k-N
i50 s51 r51
This contradicts the assumption of BIBO stability.
EXAMPLE 8.3
Determine the BIBO stability of the system with difference equations
y1 ðk 1 2Þ 1 0:1y2 ðk 1 1Þ 5 uðkÞ
y2 ðk 1 2Þ 1 0:9y1 ðk 1 1Þ 5 uðkÞ
Solution
To find the impulse response matrix, we first obtain the transfer function matrix
21
Y1 ðzÞ
z2 0:1z
5
UðzÞ
Y2 ðzÞ
0:9z z2
z2
20:1z
5
20:9z
z2
z2 ðz2 2 0:09Þ
z
UðzÞ 5
20:1
20:9
z
UðzÞ
zðz 2 0:3Þðz 1 0:3Þ
Inverse z-transforming the matrix gives the impulse response
"
GðkÞ 5
5:556fð20:3Þk 1 ð0:3Þk g
1:111δðk 2 1Þ 1 1:852fð20:3Þk 2 ð0:3Þk g
10δðk 2 1Þ 1 16:667fð20:3Þk 2 ð0:3Þk g
5:556fð20:3Þk 1 ð0:3Þk g
#
The entries of the impulse response matrix are all absolutely summable because
N
X
k50
ð0:3Þk 5
1
5 ð0:7Þ21 , N
1 2 0:3
Hence, the impulse response satisfies condition (8.10), and the system is BIBO stable.
Although it is possible to test the impulse response for BIBO stability, it is
much easier to develop tests based on the transfer function. Theorem 8.3 relates
BIBO stability to asymptotic stability and sets the stage for such a test.
THEOREM 8.3
If a discrete-time linear system is asymptotically stable, then it is BIBO stable. Furthermore,
in the absence of unstable pole-zero cancellation, the system is asymptotically stable if it is
BIBO stable.
8.2 Controllability and stabilizability
PROOF
Substituting from 7.96 into (8.10) gives
N
N X
n
X
X
:GðkÞ: 5 :D: 1
CZj Bλkj k50
k50 j51
# :D: 1
N X
n
X
:CZj B:jλj jk
k50 j51
For an asymptotically stable system, all poles are inside the unit circle. Therefore,
N
X
k50
:GðkÞ: # :D: 1
n
X
:CZj B:
j51
1
1 2 jλj j
which is finite provided that the entries of the matrices are finite. Thus, whenever the
product CZjB is nonzero for all j, BIBO and asymptotic stability are equivalent. However,
the system is BIBO stable but not asymptotically stable if the product is zero for some λj
with magnitude greater than unity. In other words, a system with unstable input-decoupling,
output-decoupling, or input-output-decoupling zeros and all other modes stable is BIBO
stable but not asymptotically stable. In the absence of unstable pole-zero cancellation,
BIBO stability and asymptotic stability are equivalent.
Using Theorem 8.3, we can test BIBO stability by examining the poles of the
transfer function matrix. Recall that, in the absence of pole-zero cancellation,
the eigenvalues of the state matrix are the system poles. If the poles are all inside the
unit circle, we conclude that the system is BIBO stable. However, we cannot conclude asymptotic stability except in the absence of unstable pole-zero cancellation.
We reexamine this issue later in this chapter.
EXAMPLE 8.4
Test the BIBO stability of the transfer function of the system of Example 8.3.
Solution
The transfer function is
z
20:1
20:9
z
GðzÞ 5
zðz 2 0:3Þðz 1 0:3Þ
The system has poles at the origin, 0.3, and 20.3, all of which are inside the unit circle.
Hence, the system is BIBO stable.
8.2 Controllability and stabilizability
When obtaining z-transfer functions from discrete-time statespace equations,
we discover that it is possible to have modes that do not affect the input-output
relationship. This occurs as a result of pole-zero cancellation in the transfer
301
302
CHAPTER 8 Properties of StateSpace Models
function. We now examine this phenomenon more closely to assess its effect on
control systems.
The time response of a system is given by a sum of the system modes weighted
by eigenvectors. It is therefore important that each system mode be influenced by
the input so as to be able to select a time response that meets design specifications.
Systems possessing this property are said to be completely controllable or simply
controllable. Modes that cannot be influenced by the control are called uncontrollable modes. A more precise definition of controllability is given next.
DEFINITION 8.4: CONTROLLABILITY
A system is said to be controllable if for any initial state x(k0) there exists a control sequence
u(k), k 5 k0, k0 1 1, . . . , kf 2 1, such that an arbitrary final state x(kf) can be reached in finite kf.
Controllability can be given a geometric interpretation if we consider a secondorder system. Figure 8.1 shows the decomposition of the state plane into a controllable subspace and an uncontrollable subspace. The controllable subspace represents
states that are reachable using the control input, whereas the uncontrollable subspace
represents states that are not reachable. Some authors prefer to use the term reachability instead of “controllability.” Reachable systems are ones where every point
in state space is reachable from the origin.1 This definition is equivalent to our
definition of controllability.
Theorem 8.4 establishes that the previous controllability definition is equivalent
to the ability of the system input to influence all its modes.
THEOREM 8.4: CONTROLLABILITY CONDITION
A linear time-invariant system is completely controllable if and only if the products
wTi Bd ; i 5 1; 2; . . . ; n are all nonzero where wTi is the ith left eigenvector of the state matrix.
Furthermore, modes for which the product wTi Bd is zero are uncontrollable.
x2
Controllable
Subspace
Uncontrollable
Subspace
u
x1
FIGURE 8.1
Controllable and uncontrollable subspaces.
1
Traditionally, controllable systems are defined as systems where the origin can be reached from
every point in the state space. This is equivalent to our definition for continuous-time systems but
not for discrete-time systems. For simplicity, we adopt the more practically relevant Definition 8.4.
8.2 Controllability and stabilizability
PROOF
Necessity and Uncontrollable Modes
Definition 8.4 with finite final time kf guarantees that all system modes are influenced by the
control. To see this, we examine the zero-input response
n
X
xZI ðkÞ 5
Zi xðk0 Þλki
i51
For any eigenvalue λi inside the unit circle, the corresponding mode decays to zero
exponentially regardless of the influence of the input. However, to go to the zero state in
finite time, it is necessary that the input influence all modes. For a zero product ZiBd, the ith
mode is not influenced by the control and can only go to zero asymptotically. Therefore,
the ith mode is uncontrollable for zero ZiBd. In Section 7.4.3, we showed that Zi is a matrix
of rank 1 given by the product of the ith right eigenvector vi of the state matrix, and its ith
left eigenvector is wi. Hence, the product ZiBd is given by
2
vi1
3
2
vi1 wTi
3
7
6 7
6
6 vi2 7
6 vi2 wT 7
i 7
6 7
6
Zi Bd 5 vi wTi Bd 56 7wTi Bd 56
7Bd
6 ^ 7
6 ^ 7
5
4 5
4
vin
vin wTi
where vij, j 5 1, . . . , n are the entries of the ith eigenvector of the state matrix. Therefore, the
product ZiBd is zero if and only if the product wTi Bd is zero.
Sufficiency
We examine the total response for any initial and terminal state. From (7.91), we have
x 5 xðkÞ 2 Akd xðk0 Þ 5
k21
X
Ak2i21
Bd uðiÞ
d
(8.11)
i5k0
Using the Cayley-Hamilton theorem, it can be shown that for k . n, no “new” terms are added
to the preceding summation. By that we mean that the additional terms will be linearly
dependent and are therefore superfluous. Thus, if the vector x cannot be obtained in n sampling
periods by proper control sequence selection, it cannot be obtained over a longer duration.
We therefore assume that k 5 n and use the expansion of (7.89) to obtain
"
#
n21 X
n
X
x5
Zj λj n2i21 Bd uðiÞ
(8.12)
i50
j51
j51
5 Lu
Zj Bd λj n21
n
X
Zj Bd λj n22
j51
. . .:
n
X
Zj Bd λj
j51
n
X
x5
"
The outer summation in (8.12) can be written in matrix form to obtain
2
3
uð0Þ
6 7
7
6
7
6
uð1Þ
7
6
#
6 7
n
X
7
6
:
7
6
Zj Bd 6
7
6 7
j51
7
6
6 uðn 2 2Þ 7
7
6
4 5
uðn 2 1Þ
(8.13)
303
CHAPTER 8 Properties of StateSpace Models
The matrix Bd is n 3 m and is assumed of full rank m. We need to show that the n 3 m.n
matrix L has full row rank. The constituent matrices Zi, i 5 1, 2, . . . , n are of rank 1, and each
has a column linearity independent of the other constituent matrices. Therefore, provided
that the individual products ZiBd are all nonzero, each submatrix of L has rank 1, and L has n
linearly independent columns (i.e., L has full rank n). Because the rank of a matrix cannot
exceed the number of rows for fewer rows than columns, we have
rankfLjxg 5 rankfLg 5 n
(8.14)
This condition guarantees the existence of a solution u to (8.12). However, the solution
is, in general, nonunique because the number of elements of u (unknowns) is m.n . n.
A unique solution exists only in the single-input case (m 5 1). It follows that if the products
ZiBd, i 5 1, 2, . . . , n are all nonzero, then one can solve for the vector of input sequences
needed to reach any x(kf) from any x(k0). As discussed in the proof of necessity, the product
ZiBd is zero if and only if the product wTi Bd is zero.
Theorem 8.5 gives a simpler controllability test, but, as first stated, it does not
provide a means of determining which of the system modes is uncontrollable.
THEOREM 8.5: CONTROLLABILITY RANK CONDITION
Ad Bd
C 5 Bd
A linear time-invariant system is completely controllable if and only if the n 3 m.n controllability matrix
. . . An21
d Bd
(8.15)
has rank n.
PROOF
5 Cu
We first write (8.13) in the form (recall the Cayley-Hamilton theorem)
3
2
uðn 2 1Þ
7
6
6 7
7
6
6 uðn 2 2Þ 7
7
6
7
6
7
6
x 5 Bd Ad Bd . . . An21
:
B
d
d
7
6
6 7
7
6
7
6
uð1Þ
7
6
4 5
uð0Þ
304
(8.16)
We now have a system of linear equations for which (8.15) is a necessary and sufficient condition for a solution u to exist for any x. Hence, (8.15) is a necessary and sufficient condition
for controllability.
If the controllability matrix C is rank deficient, then the rank deficiency is
equal to the number of linearly independent row vectors that are mapped to zero
on multiplication by C. These row vectors are the uncontrollable states of the
8.2 Controllability and stabilizability
system as well as the left eigenvectors of the state matrix A. The vectors are also
the transpose of the right eigenvectors of AT. Thus, the rank test can be used to
determine the uncontrollable modes of the system if the eigenstructure of AT is
known.
EXAMPLE 8.5
Determine the controllability of the following state equation:
32
2
3 2 22
3 2
22
0
x1 ðkÞ
1
x1 ðk 1 1Þ
76
0
0
1
6
6
7 6
74 x2 ðkÞ 7
1
0
4
4 x2 ðk 1 1Þ 5 5 6
5
4 5
1
x3 ðk 1 1Þ
x
ðkÞ
3
0
20:4 20:5
0
3
7
1 5uðkÞ
1
Solution
1
6
56
0
4
1
0
1
1
22
22
1
1
20:5
20:9
2
The controllability matrix for the system is
h
i
C 5 Bd Ad Bd A2d Bd
2
2
3
20:5
7
20:9 7
5
20:15
0:05
The matrix has rank 3, which implies that the third-order system is controllable.
In fact, the first three columns of the matrix are linearly independent, and the same
conclusion can be reached without calculating the entire controllability matrix. In general,
one can gradually compute more columns until n linearly independent columns are
obtained to conclude controllability for an nth-order system.
Although the controllability tests are given here for discrete-time systems,
they are applicable to continuous-time systems. This fact is used in the following
two examples.
EXAMPLE 8.6
Show that the following state equation is uncontrollable, and determine the uncontrollable
mode. Then obtain the discrete-time state equation, and verify that it is also uncontrollable.
x_1 ðtÞ
0 1 x1 ðtÞ
21
5
1
uðtÞ
x_2 ðtÞ
0 0 x2 ðtÞ
0
Solution
The system has a zero eigenvalue with multiplicity 2. The controllability matrix for the
system is
h
i
C 5 B AB
2
3
21 0
5
54
0
0
305
CHAPTER 8 Properties of StateSpace Models
The matrix has rank 1 (less than 2), which implies that the second-order system has one
(2 2 1) uncontrollable mode. In fact, integrating the state equation reveals that the second
state variable is governed by the equation
x2 ðtÞ 5 x2 ðt0 Þ;
t $ t0
Therefore, the second state variable cannot be altered using the control and corresponds
to an uncontrollable mode. The first state variable is influenced by the input and can be
arbitrarily controlled.
The state-transition matrix for this system can be obtained directly using the series
expansion because the matrix Ai is zero for i . 1. Thus, we have the discrete-time state
matrix
1 T
Ad 5 eAT 5 I2 1 AT 5
0 1
The discrete-time input matrix is
ðT
ðT
21
½I2 1 Aτ Bd 5
eAτ Bdτ 5
dτ 5
0
0
0
2T
0
The controllability matrix for the discrete-time system is
h
i
C 5 Bd Ad Bd
2T 2T
5
0
0
306
As with the continuous time system, the matrix has rank 1 (less than 2), which implies
that the second-order system is uncontrollable.
The solution of the difference equation for the second state variable gives
x2 ðkÞ 5 x2 ðk0 Þ;
k $ k0
with initial time k0. Thus, the second state variable corresponds to an uncontrollable unity
mode (e0), whereas the first state variable is controllable as with the continuous system.
EXAMPLE 8.7
Show that the state equation for the motor system
2 3 2
32 3 2 3
x_1
0
1
0
x1
1
4 x_2 5 5 4 0
0
1 54 x2 5 1 4 0 5u
x_3
x3
0 210 211
0
with DAC and ADC is uncontrollable, and determine the uncontrollable modes. Obtain
the transfer functions of the continuous-time and discrete-time systems, and relate the
uncontrollable modes to the poles of the transfer function.
Solution
From Example 7.7 the state-transition matrix of the system is
2
3
2
2
3
0 210 21 2t
0
10 11 1 0
e
e
1 5
140
eAt 5 4 0 0 0 5 1 4 0 10
10
9
0 210 21
0
0 0 0
1
210
100
3
1
e210t
210 5
90
100
8.2 Controllability and stabilizability
Multiplying by the input matrix, we obtain
2 3
2 3
2 3
2 3
0 210t
1
1
0 2t
e
e
e0
0
4
5
4
5
4
5
1 0
5405
e B5 0 e 1 0
9
90
10
0
0
0
0
At
Thus, the modes e2t and e210t are both uncontrollable for the analog subsystem. The two
modes are also uncontrollable for the digital system, including DAC and ADC.
Using the results of Example 7.9, the input matrix for the discrete-time system is
2T
Bd 5 Z1 BT 1 Z2 Bð1 2 e
210T
Þ 1 Z3 Bð1 2 e
2 3
1
Þ 5 4 0 5T
0
and the matrix Ad 5 eAT. The controllability matrix for the discrete-time system is
h
C 5 Bd Ad Bd
2
1 1 1
6
6
540 0 0
0 0 0
A2d Bd
3
i
7
7T
5
which clearly has rank 1, indicating that there are 2 (i.e., 3 2 1) uncontrollable modes. The
left eigenvectors corresponding to the eigenvalues e2T and e210T have zero first entry, and
their product with the controllability matrix C is zero. Hence, the two corresponding modes
are not controllable. The zero eigenvalue has the left eigenvector [1, 0, 0], and its product
with C is nonzero. Thus, the corresponding mode is controllable.
The transfer function of the continuous-time system with output x1 is
GðsÞ 5 C½sI3 2A21 B 5
1
s
This does not include the uncontrollable modes (input-decoupling modes) because they
cancel when the resolvent matrix and the input matrix are multiplied.
The z-transfer function corresponding to the system with DAC and ADC is
GZAS ðzÞ 5
z21
GðsÞ
T
Z L 21
5
z
s
z21
The reader can easily verify that the transfer function is identical to that obtained using
the discrete-time statespace representation. It includes the pole at e0 but has no poles at
e2T or e210T.
8.2.1 MATLAB commands for controllability testing
The MATLAB commands to calculate the controllability matrix and determine its
rank are
.. c 5 ctrbðA;CÞ
.. rankðcÞ
307
CHAPTER 8 Properties of StateSpace Models
8.2.2 Controllability of systems in normal form
Checking controllability is particularly simple if the system is given in normal
form—that is, if the state matrix is in the form
A 5 diagfλ1 ðAÞ; λ2 ðAÞ; . . . ; λn ðAÞg
The corresponding state-transition matrix is
eAt 5 L f½sIn 2A21 g 5 diagfeλ1 ðAÞt ; eλ2 ðAÞt ; . . . ; eλn ðAÞt g
(8.17)
The discrete-time state matrix is
Ad 5 eAT 5 diagfeλ1 ðAÞT ; eλ2 ðAÞT ; . . . ; eλn ðAÞT g
(8.18)
5 diagfλ1 ; λ2 ; . . . ; λn g
Theorem 8.6 establishes controllability conditions for a system in normal form.
THEOREM 8.6: CONTROLLABILITY OF SYSTEMS IN NORMAL FORM
A system in normal form is controllable if and only if its input matrix has no zero rows.
Furthermore, if the input matrix has a zero row, then the corresponding mode is uncontrollable.
PROOF
Necessity and Uncontrollable Modes
For a system in normal form, the state equations are in the form
xi ðk 1 1Þ 5λi xi ðkÞ 1 bTi uðkÞ;
i 5 1; 2; . . . ; n
where bTi is the ith row of the input matrix Bd. For a zero row, the system is unforced and can
only converge to zero asymptotically for jλij inside the unit circle. Because controllability
requires convergence to any final state (including the origin) in finite time, the ith mode is
uncontrollable for zero bTi .
Sufficiency
From the sufficiency proof of Theorem 8.4, we obtain equation (8.13) to be solved for
the vector u of controls over the period k 5 0,1, . . . ,n 2 1. The solution exists if the matrix
L in (8.13) has rank n. For a system in normal form, the state-transition matrix is in the
form
Akd 5 diagfλki g
. . .:
diagfλj gBd
diagfλj n22 gBd
L 5 diagfλj n21 gBd
and the n 3 n.m matrix L is in the form
308
Bd
8.2 Controllability and stabilizability
Substituting for Bd in terms of its rows and using the rules for multiplying partitioned
matrices, we obtain
2
λ1 n21 bT1
6 n21 T
6 λ2 b
2
6
L56
6
^
4
λn n21 bTn
T
λn22
1c b1
λ2 n22 bT2
^
λn n22 bTn
?
bT1
3
7
bT2 7
7
7
& ^ 7
5
?
?
bTn
For a matrix Bd with no zero rows, the rows of L are linearly independent and the matrix has
full rank n. This guarantees the existence of a solution u to (8.13).
EXAMPLE 8.8
Determine the controllability of the system
32
3 2
22
0
0
x1 ðkÞ
1
7
6
0
0
0
6
7
7 6
76
4 x2 ðk 1 1Þ 5 5 6
4 54 x2 ðkÞ 5 1 4 0
1
x3 ðk 1 1Þ
x3 ðkÞ
0
0
20:5
2
x1 ðk 1 1Þ
3
2
0
3
7
1 5uðkÞ
1
Solution
The system is in normal form, and its input matrix has no zero rows. Hence, the system is
completely controllable.
8.2.3 Stabilizability
The system in Example 8.8 is controllable but has one unstable eigenvalue
(outside the unit circle) at (22). This clearly demonstrates that controllability
implies the ability to control the modes of the system regardless of their
stability. If the system has uncontrollable modes, then these modes cannot be
influenced by the control and may or may not decay to zero asymptotically.
Combining the concepts of stable and controllable modes gives the following
definition.
DEFINITION 8.5: STABILIZABILITY
A system is said to be stabilizable if all its uncontrollable modes are asymptotically stable.
Physical systems are often stabilizable rather than controllable. This poses no
problem provided that the uncontrollable dynamics decay to zero sufficiently fast
so as not to excessively slow down the system.
309
CHAPTER 8 Properties of StateSpace Models
EXAMPLE 8.9
Determine the controllability and stabilizability of the system
32
2
3 2 22
3 2
3
0
0
x1 ðkÞ
x1 ðk 1 1Þ
1 0
7
6
0
0
0
6
7
7 6
7
76
4 x2 ðk 1 1Þ 5 5 6
4 54 x2 ðkÞ 5 1 4 0 0 5uðkÞ
1 1
x3 ðk 1 1Þ
x3 ðkÞ
0
0
20:5
Solution
The system is in normal form, and its input matrix has one zero row corresponding to its
zero eigenvalue. Hence, the system is uncontrollable. However, the uncontrollable mode at
the origin is asymptotically stable, and the system is therefore stabilizable.
An alternative procedure to determine the stabilizability of the system is to
transform it to a form that partitions the state into a controllable part and an
uncontrollable part, and then determine the asymptotic stability of the uncontrollable part. We can do this with Theorem 8.7.
THEOREM 8.7: STANDARD FORM FOR UNCONTROLLABLE SYSTEMS
Consider the pair (Ad, Bd) with nc controllable modes and n 2 nc uncontrollable modes and
the transformation matrix
321
2
qT1
7
6
6 ^ 7
7
6
6 qT 7 2
321
6 nc 7
Q1
7
6
75 4 5
(8.19)
Tc 5 6
7
6
Q2
7
6 qT
6 nc 11 7
7
6
7
6
4 ^ 5
qTn
where fqTi ; i 5 nc 1 1; . . . ; ng are linearly independent vectors in the null space of the controllability matrix C and fqTi ; i 5 1; . . . ; nc g are arbitrary linearly independent vectors selected to
make Tc nonsingular.
The transformed state-transition matrix A and the transformed input matrix B have the
following form:
2
3
2
3
Acuc
Ac
Bc
A 5 4 5 B 5 4 5
(8.20)
0
Auc
0
310
where Ac is an nc 3 nc matrix, Bc is an nc 3 m matrix, and the pair (Ac, Bc) is controllable.
PROOF
If the rank of the controllability matrix C is nc, there exist linearly independent vectors
fqTi ; i 5 nc 1 1; . . . ; ng satisfying qTi C 5 0T , and the transformation Tc can be obtained by
adding nc linearly independent vectors. The null space of the controllability matrix is the
8.2 Controllability and stabilizability
unreachable space of the system and is also spanned by the set of left eigenvectors
corresponding to uncontrollable modes. Hence, the vectors fqTi ; i 5 nc 1 1; . . . ; ng are linear
combinations of the left eigenvectors of the matrix Ad corresponding to uncontrollable modes.
We can therefore write the transformation matrix in the form
2
321
Q1
4
5
Tc 5
Q2 Wuc
where the matrix of coordinates Q2 is nonsingular and Wuc is a matrix whose rows are the left
eigenvectors of Ad,fwTi ; i 5 1; . . . ; n 2 nc g, corresponding to uncontrollable modes. By the
eigenvector controllability test, if the ith mode is uncontrollable, we have
wTi Bd 5 0T ; i 5 1; . . . ; n 2 nc
Therefore, transforming the input matrix gives
2
B 5 Tc21 Bd
3
2
3
Q1
Bc
4
5
4
5
5
Bd 5
0
Q2 Wuc
Next, we show that the transformation yields a state matrix in the form (8.20). We first
recall that the left and right eigenvectors can be scaled to satisfy the condition
wTi vi 5 1;
i 5 1; . . . ; n 2 nc
If we combine the condition for all the eigenvectors corresponding to uncontrollable modes,
we have
Wuc Vuc 5 ½wTi vj 5 In2nc
Tc 5 R1
We write the transformation matrix in the form
R2 5 R1
Vuc R2
which satisfies the condition
i
Vuc R2
2
3
0
Inc
5 4 5
0
In2nc
3
Q1
h
5 4 5 R1
Q2 Wuc
2
Tc21 Tc
Because Q2 is nonsingular, we have the equality
Wuc R1 5 0
The similarity transformation gives
2
3
Q1
Tc21 Ad Tc 5 4 5Ad R1 Vuc R2
Q2 Wuc
2
3
Q1 Ad R1
Q1 Ad Vuc R2
4
5 5
Q2 Wuc Ad R1 Q2 Wuc Ad Vuc R2
(8.21)
311
CHAPTER 8 Properties of StateSpace Models
Because premultiplying the matrix Ad by a left eigenvector gives the same eigenvector scaled
by an eigenvalue, we have the condition
Wuc Ad 5 Λu Wuc ; Λu 5 diagfλu1 ; . . . ; λu;n2nc g
3 2
3
Acuc
Ac
Q1 Ad Vuc R2
Q1 Ad R1
4
5
4
5
5 5
0
Auc
0
Q2 Λu R2
2
with λu denoting an uncontrollable eigenvalue. This simplifies our matrix equality (8.21) to
2
3 2
3
Q1 Ad Vuc R2
Q1 Ad Vuc R2
Q1 Ad R1
Q1 Ad R1
21
Tc Ad Tc 5 4 5 5 4 5
Q2 Λu Wuc R1 Q2 Λu Wuc Vuc R2
Q2 Λu Wuc R1
Q2 Λu R2
Because all the uncontrollable modes correspond to the eigenvalues of the matrix Auc, the
remaining modes are all controllable, and the pair (Ac, Bc) is completely controllable.
Because the pair (Ad, Bd) has nc controllable modes, the rank of the controllability matrix is nc and the transformation matrix Tc is guaranteed to exist. The columns fqTi ; i 5 nc 1 1; . . . ; ng of the inverse transformation matrix have a geometric
interpretation. They form a basis set for the unobservable subspace of the system
and can be selected as the eigenvectors of the uncontrollable modes of the system.
The remaining columns of Tc form a basis set of the controllable subspace and can
be conveniently selected as linearly independent columns fqTi ; i 5 i; . . . ; ng of the
controllability matrix. Clearly, these vectors are not in the null space of the controllability matrix as they satisfy the equation qTi ; C 6¼ 0, i 5 1, . . . , n. Starting with
these vectors, the transformation matrix can be formed by selecting nnc additional linearly independent vectors.
The similarity transformation is a change of basis that allows us to separate
the controllable subspace from the uncontrollable subspace. After transformation,
we can check the stabilizability of the system by verifying that all the eigenvalues
of the matrix Auc are inside the unit circle.
EXAMPLE 8.10
Determine the controllability, stability, and stabilizability of the system
2
3 2
32
3 2
3
21:85 4:2 20:15
x1 ðkÞ
4 3
x1 ðk 1 1Þ
4 x2 ðk 1 1Þ 5 5 4 20:3 0:1
5
4
5
4
x2 ðkÞ 1 1 1 5uðkÞ
0:3
x3 ðk 1 1Þ
x3 ðkÞ
21:35 4:2 20:65
2 1
Solution
We first find the controllability matrix of the system
h
i
C 5 Bd Ad Bd A2d Bd
2
4 3
23:5 21:5
6
6
561 1
20:5 20:5
4
2 1
22:5 20:5
312
4:75 0:75
3
7
7
0:25 0:25 7
5
4:25 0:25
8.3 Observability and detectability
2
4 3
6
Tc 5 6
41 1
2 1
The matrix has rank 2, and the first two columns are linearly independent. The dimension
of the null space is 3 2 2 5 1. We form a transformation matrix of the two first columns of
the controllability matrix and a third linearly independent column, giving
0
3
7
07
5
1
The system is transformed to
x1 ðk 1 1Þ
3
2
3
22 0
21:05 2
3 21 0 3
x1 ðkÞ
7
6 1:5 20:5
1:35 76
6
7 6
7 60 1 7
7
74 x2 ðkÞ 5 1 6
4 x2 ðk 1 1Þ 5 5 6
4 5uðkÞ
4 5
ðkÞ
x3 ðk 1 1Þ
x
3
0
0
0:1
0 0
2
The system has one uncontrollable mode corresponding to the eigenvalue at 0.1,
which is inside the unit circle. The system is therefore stabilizable but not controllable.
The two other eigenvalues are at 22, outside the unit circle, and at 20.5, inside the unit
circle, and both corresponding modes are controllable. The system is unstable because
one eigenvalue is outside the unit circle.
8.3 Observability and detectability
To effectively control a system, it may be advantageous to use the system state to
select the appropriate control action. Typically, the output or measurement vector
for a system includes a combination of some but not all the state variables. The
system state must then be estimated from the time history of the measurements
and controls. However, state estimation is only possible with proper choice of the
measurement vector. Systems whose measurements are so chosen are said to be
completely observable or simply observable. Modes that cannot be detected
using the measurements and the control are called unobservable modes. A more
precise definition of observability is given next.
DEFINITION 8.6: OBSERVABILITY
A system is said to be observable if any initial state x(k0) can be estimated from the control
sequence u(k), k 5 k0,k0 1 1, . . . , kf 2 1 and the measurements y(k), k 5 k0, k0 1 1, . . . , kf.
As in the case of controllability, observability can be given a geometric interpretation if we consider a second-order system. Figure 8.2 shows the decomposition of the state plane into an observable subspace and an unobservable subspace.
The observable subspace includes all initial states that can be identified using the
measurement history, whereas the unobservable subspace includes states that are
indistinguishable.
313
314
CHAPTER 8 Properties of StateSpace Models
x2
Observable
Subspace
Unobservable
Subspace
y
x1
FIGURE 8.2
Observable and unobservable subspaces.
The following theorem establishes that this observability definition is equivalent
to the ability to estimate all system modes using its controls and measurements.
THEOREM 8.8: OBSERVABILITY
A system is observable if and only if Cvi is nonzero for i 5 1, 2, . . . , n, where vi is the ith
eigenvector of the state matrix. Furthermore, if the product C vi is zero, then the ith mode is
unobservable.
PROOF
We first define the vector
yðkÞ 5 yðkÞ 2 C
k21
X
Ak2i21
Bd uðiÞ 2 DuðkÞ
d
i5k0
5 CAkd xðk0 Þ 5
n
X
CZi λki xðk0 Þ
i51
Sufficiency
Stack the output vectors yðiÞ; i 5 1; . . . ; k to obtain
3
2 X
n
3
2
CZ
j
7
6
yð0Þ
7
6
7
6 7 6 nj51
7
7 6 X
6
7
7 6
6
yð1Þ
CZ
λ
j j 7
7 6
6
7
6 7 6 j51
7
7 6
6
7xðk0 Þ
756
6
^
^
7
7 6 n
6
7
6 7 6X
7 6
6
CZj λj n22 7
7
6 yðn 2 2Þ 7 6
7
7 6 j51
6
7
4 5 6
n
7
6X
4
yðn 2 1Þ
n21 5
CZj λj
j51
5 Lxðk0 Þ
The products CZi are
CZi 5 Cvi wTi 5 Cvi wi1
5 C wi1 vi
wi2 vi
wi2
...
. . . win vi
win
(8.22)
8.3 Observability and detectability
where wij, j 5 1, . . . , n are the entries of the ith left eigenvector, and vi is the ith right eigenvector of the state matrix A (see Section 7.4.3). Therefore, if the products are nonzero, the
matrix L has n linearly independent columns (i.e., has rank n). An n 3 n submatrix of L can
be formed by discarding the dependent rows. This leaves us with n equations in the n
unknown entries of the initial condition vector. A unique solution can thus be obtained.
Necessity
Let x(k0) be in the direction of vi, the ith eigenvector of Ad, and let CZi be zero. Then the
vector y(k) is zero for any k regardless of the amplitude of the initial condition vector (recall
that Zivj 5 0 whenever i6¼j). Thus, all initial vectors in the direction vi are indistinguishable
using the output measurements, and the system is not observable.
Theorem 8.9 establishes an observability rank test that can be directly checked
using the state and output matrices.
THEOREM 8.9: OBSERVABILITY RANK CONDITION
A linear time-invariant system is completely observable if and only if the l.n 3 n observability
matrix
3
C
6 7
7
6
6 CAd 7
7
6
6
O 56 7
7
6 ^ 7
7
6
4 5
n21
CAd
2
(8.23)
has rank n.
PROOF
We first write (8.22) in the form
3
yð0Þ
3
2
6 7
C
7
6
6 yð1Þ 7 6 7
7 6
7
6
6 7 6 CAd 7
7 6
7
6
7 5 6 7xðk0 Þ
6
^
7 6
7
6
6 7 6 ^ 7
7 6
7
6
6 yðn 2 2Þ 7 4 5
7
6
4 5
CAn21
d
yðn 2 1Þ
2
(8.24)
5 Oxðk0 Þ
We now have a system of linear equations that can include, at most, n linearly independent
equations. If n independent equations exist, their coefficients can be used to form an n 3 n
invertible matrix. The rank condition (8.23) is a necessary and sufficient condition for a
unique solution x(k0) to exist. Thus, (8.23) is a necessary and sufficient condition for
observability.
315
316
CHAPTER 8 Properties of StateSpace Models
If the observability matrix O is rank deficient, then the rank deficiency is
equal to the number of linearly independent column vectors that are mapped to
zero on multiplication by O. This number equals the number of unobservable
states of the system, and the column vectors mapped to zero are the right eigenvectors of the state matrix. Thus, the rank test can be used to determine the unobservable modes of the system if the eigenstructure of the state matrix is known.
EXAMPLE 8.11
Determine the observability of the system using two different tests:
0231 I2
C5 0 0 1
A5
0 23 4
If the system is not completely observable, determine the unobservable modes.
Solution
Because the state matrix is in companion form, its characteristic equation is easily
obtained from its last row. The characteristic equation is
λ3 2 4λ2 1 3λ 5 λðλ 2 1Þðλ 2 3Þ 5 0
Hence, the system eigenvalues are {0, 1, 3}.
The companion form of the state matrix allows us to write the modal matrix of eigenvectors as the Van der Monde matrix:
2
3
1 1 1
V 540 1 35
0 1 9
Observability is tested using the product of the output matrix and the modal matrix:
2
3
1 1 1
CV 5 0 0 1 4 0 1 3 5 5 0 1 9
0 1 9
The product of the output matrix and the eigenvector for the zero eigenvalue is zero. We conclude that the system has an output-decoupling zero at zero (i.e., one unobservable mode).
The observability matrix of the system is
3
2
3
2
0
0
1
C
6 7
6 7
7
6
7
6
CAd 7 5 6 0 23
4 7
O56
7
6
7
6
4 5
4 5
2
CAd
0 212 13
which has rank 2 5 3 2 1. Hence, the system has one unobservable mode. The product of
the observability matrix and the eigenvector for the zero eigenvalue is zero. Hence, it corresponds to the unobservable mode of the system.
8.3.1 MATLAB commands
The MATLAB commands to test observability are
.. o 5 obsvðA; CÞ% Obtain the observability matrix
.. rankðoÞ
8.3 Observability and detectability
8.3.2 Observability of systems in normal form
The observability of a system can be easily checked if the system is in normal
form by exploiting Theorem 8.10.
THEOREM 8.10: OBSERVABILITY OF NORMAL FORM SYSTEMS
A system in normal form is observable if and only if its output matrix has no zero columns.
Furthermore, if the input matrix has a zero column, then the corresponding mode is
unobservable.
PROOF
The proof is similar to that of the Controllability Theorem 8.6 and is left as an exercise.
8.3.3 Detectability
If a system is not observable, it is preferable that the unobservable modes be stable. This property is called detectability.
DEFINITION 8.7: DETECTABILITY
A system is detectable if all its unobservable modes decay to zero asymptotically.
EXAMPLE 8.12
Determine the observability and detectability of the system
2
3
2
3
0
0 2 x1 ðkÞ 3 2 1
3
0 7
7 6
76
4 x ðkÞ 5 1 4 0
5 2
1
x3 ðkÞ
0 22
2
3
x1 ðkÞ
6
7
1 0 4 x2 ðkÞ 5
x3 ðkÞ
20:4
0
6
7 6
6
4 x2 ðk 1 1Þ 5 5 4
x3 ðk 1 1Þ
0
x1 ðk 1 1Þ
yðkÞ 5 1
0
3
7
0 5uðkÞ
1
Solution
The system is in normal form, and its output matrix has one zero column corresponding to
its eigenvalue 22. Hence, the system is unobservable. The unobservable mode at 22,j22j . 1,
is unstable, and the system is therefore not detectable.
Similar to the concepts described for stabilizability, a procedure to determine
the detectability of the system is to transform it to a form that allows partitioning
the state into an observable part and an unobservable part and then determining
the asymptotic stability of the unobservable part. We do this with Theorem 8.11.
317
CHAPTER 8 Properties of StateSpace Models
THEOREM 8.11: STANDARD FORM FOR UNOBSERVABLE SYSTEMS
h
To 5 To1
i h
To2 5 t1
. . . tn2n0
Consider the pair (Ad, Cd) with n0 observable modes and n 2 n0 unobservable modes and the
n 3 n transformation matrix
tn2n0 11
. . . tn
i
2
3
Auo
Au
h
A 5 4 5 C 5 0l 3 n2n0
0n0 3 n2n0 Ao
where {ti, i 5 1, . . . , n 2 no} are linearly independent vectors in the null space of the observability matrix O and {ti, i 5 n 2 no 1 1, . . . , n} are arbitrary vectors selected to make To invertible. The transformed state-transition matrix A and the transformed input matrix C have the
following form:
318
i
Co
where Ao is an n0 3 n0 matrix, Co is an l 3 n0 matrix, and the pair (Ao, Co) is observable.
PROOF
The proof is similar to that of Theorem 8.7, and it is left as an exercise.
Because the pair (Ad, Cd) has no observable modes, the rank of the observability
matrix is no and the transformation matrix To is guaranteed to exist. The columns
{ti, i 5 1, . . . , n 2 no} of the transformation matrix have a geometric interpretation.
They form a basis set for the unobservable subspace of the system and can be
selected as the eigenvectors of the unobservable modes of the system. The remaining
columns of To form a basis set of the observable subspace and can be selected as the
transpose of linearly independent rows rT of the observability matrix. Clearly, these
vectors are not in the null space of the observability ability matrix as they satisfy
the equation O r ¼
6 0. Starting with these vectors, the transformation matrix can be
formed by selecting nno additional linearly independent vectors.
The similarity transformation is a change of basis that allows us to separate
the observable subspace from the unobservable subspace. After transformation,
we can check the detectability of the system by verifying that all the eigenvalues
of the matrix Au are inside the unit circle.
EXAMPLE 8.13
Determine the observability and detectability of the system
2
3 2
2:0
x1 ðk 1 1Þ
6
7 6
5
21:1
ðk
1
1Þ
x
4 2
5 4
2:6
x3 ðk 1 1Þ
3 2
x1 ðkÞ
1
76
7 6
22:5 21:15 54 x2 ðkÞ 5 1 4 1
6:8
2:8
0
x3 ðkÞ
2
3
ðkÞ
x
1
2 10 3 6
7
yðkÞ 5
4 x2 ðkÞ 5
1 8 3
x3 ðkÞ
4:0
2:0
32
0
3
7
1 5uðkÞ
1
8.4 Poles and zeros of multivariable systems
Solution
We first find the observability matrix of the system
3
2
2
10
3
3 6 1
2
8
3 7
7
6
C
6 7
7
6 7 6
7 6 0:8 3:4 0:9 7
6
7
CAd 7 5 6
O 56
7 6 1:0 4:4 1:2 7
6
7
4 5 6
6 7
7
6
CA2d
4 0:2 0:82 0:21 5
0:28 0:16 0:3
The matrix has rank 2, and the system is not completely observable. The dimension of the
null space is 3 2 2 5 1, and there is only one vector satisfying O v1 5 0. It is the eigenvector
v1 corresponding to the unobservable mode and is given by
2
3
1
v1 5 420:5 5
1
The transformation matrix To can then be completed by adding any two linearly independent columns—for example,
3
2
1
1 0
7
6
7
To 5 6
4 20:5 0 1 5
1
0 0
The system is transformed to
3
1
21 5uðkÞ
1:5
3
2
2
2:6
6:8 2
3
3 2
x1 ðk 1 1Þ
0
6 7 x1 ðkÞ
74
4 x2 ðk 1 1Þ 5 5 6
6 0 20:6 22:8 7 x2 ðkÞ 5 1 4 1
5
4
x3 ðk 1 1Þ
x3 ðkÞ
1
0
0:2
0:9
3
"
#2
x1 ðkÞ
0 2 10
4
yðkÞ 5
x2 ðkÞ 5
0 1 8
x3 ðkÞ
2
The system has one unobservable mode corresponding to the eigenvalue at 2, which is outside the unit circle. The system is therefore not detectable.
8.4 Poles and zeros of multivariable systems
As for single-input-single-output (SISO) systems, poles and zeros determine the
stability, controllability, and observability of a multivariable system. For SISO
systems, zeros are the zeros of the numerator polynomial of the scalar transfer
function, whereas poles are the zeros of the denominator polynomial. For multiinput-multi-output (MIMO) systems, the transfer function is not scalar, and it is
no longer sufficient to determine the zeros and poles of individual entries of
the transfer function matrix. In fact, element zeros do not play a major role in
319
320
CHAPTER 8 Properties of StateSpace Models
characterizing multivariable systems and their properties beyond their effect on the
shape of the time response of the system. In our discussion of transfer function
matrices in Section 7.8, we mention three important types of multivariable zeros:
1. Input-decoupling zeros, which correspond to uncontrollable modes
2. Output-decoupling zeros, which correspond to unobservable modes
3. Input-output-decoupling zeros (which correspond to modes that are neither
controllable nor observable)
We can obtain those zeros and others, as well as the system poles, from the transfer function matrix.
8.4.1 Poles and zeros from the transfer function matrix
We first define system poles and zeros based on a transfer function matrix of a
MIMO system.
DEFINITION 8.8: POLES
Poles are the roots of the least common denominator of all nonzero minors of all orders of
the transfer function matrix.
The least common denominator of the preceding definition is known as the pole
polynomial and is essential for the determination of the poles. The pole polynomial
is in general not equal to the least common denominator of all nonzero elements,
known as the minimal polynomial. The MATLAB command pole gives the poles
of the system based on its realization. Unfortunately, the command may not use the
minimal realization of the transfer function of the system. Thus, the command may
give values for the pole that need not be included in a minimal realization of the
transfer function.
From the definition, the reader may guess that the poles of the system are the
same as the poles of the elements. Although this is correct, it is not possible to guess
the multiplicity of the poles by inspection, as the following example demonstrates.
EXAMPLE 8.14
Determine the poles of the transfer function matrix
3
2
1
1
6
z21
ðz 2 1Þðz 2 0:5Þ 7
7
6
7
GðzÞ 5 6
7
6
z
2
0:1
1
5
4
ðz 2 0:2Þðz 2 0:5Þ
z 2 0:2
Solution
The least common denominator of the matrix entries is
ðz 2 0:2Þðz 2 0:5Þðz 2 1Þ
8.4 Poles and zeros of multivariable systems
U1(z)
1
z − 0.5
+
z − 0.1
z − 0.5
+
Y1(z)
1
z−1
1
z − 0.2
Y2(z)
U2(z)
FIGURE 8.3
Block diagram of the system described in Example 8.14.
The determinant of the matrix is
det½GðzÞ 5
1
z 2 0:1
2
ðz 2 1Þðz 2 0:2Þ ðz 2 0:2Þðz20:5Þ2 ðz 2 1Þ
The denominator of the determinant of the matrix is
ðz 2 0:2Þðz20:5Þ2 ðz 2 1Þ
The least common denominator of all the minors is
ðz 2 0:2Þðz20:5Þ2 ðz 2 1Þ
Thus, the system has poles at {0.2, 0.5, 0.5, 1}. Note that some poles of more than one
element of the transfer function are not repeated poles of the system.
For this relatively simple transfer function, we can obtain the minimal number of
blocks needed for realization, as shown in Figure 8.3. This clearly confirms our earlier
determination of the system poles.
The MATLAB command pole gives
.. poleðgÞ
ans 5
1:0000
0:5000
0:2000
1:0000
0:5000
0:2000
This includes two poles that are not needed for the minimal realization of the transfer
function.
The definition of the multivariable zero is more complex than that of the pole,
and several types of zeros can be defined. We define zeros as follows.
321
322
CHAPTER 8 Properties of StateSpace Models
DEFINITION 8.9: SYSTEM ZEROS
Consider an l 3 m transfer function matrix written such that each entry has a denominator
equal to the pole polynomial. Zeros are values that make the transfer function matrix rank
deficient—that is, values z0 that satisfy either of the two equations
Gðz0 Þwin 5 0; win 6¼ 0
(8.25)
wTout Gðz0 Þ 5 0; wout 6¼ 0
(8.26)
for some nonzero real vector win known as the input zero direction or a nonzero real vector
wout known as the output zero direction.
For any matrix, the rank is equal to the order of the largest nonzero minor.
Thus, provided that the terms have the appropriate denominator, the zeros are the
divisors of all minors of order equal to the minimum of the pair (l, m). For a
square matrix, zeros of the determinant rewritten with the pole polynomial as its
denominator are the zeros of the system. These definitions are examined in the
following examples.
EXAMPLE 8.15
Determine the z-transfer function of a digitally controlled single-axis milling machine with
a sampling period of 40 ms if its analog transfer function is given by2
GðsÞ 5 diag
3150
1092
;
ðs 1 35Þðs 1 150Þ ðs 1 35Þðs 1 30Þ
Find the poles and zeros of the transfer function.
Solution
Using the MATLAB command c2d, we obtain the transfer function matrix
GZAS ðzÞ 5 diag
0:4075ðz 1 0:1067Þ
0:38607ðz 1 0:4182Þ
;
ðz 2 0:2466Þðz 2 0:2479 3 1022 Þ ðz 2 0:2466Þðz 2 0:3012Þ
Because the transfer function is diagonal, the determinant is the product of its diagonal
terms. The least common denominator of all nonzero minors is the denominator of the
determinant of the transfer function
ðz 2 0:2479 3 1022 Þðz20:2466Þ2 ðz 2 0:3012Þ
We therefore have poles at {0.2479 3 1022, 0.2466, 0.2466, 0.3012}. The system is stable
because all the poles are inside the unit circle.
To obtain the zeros of the system, we rewrite the transfer function in the form
GZAS ðzÞ 5
3
ðz 2 0:2466Þ
ðz20:2466Þ2 ðz 2 0:2479 3 1022 Þðz 2 0:3012Þ
0:4075ðz 1 0:1067Þðz 2 0:3012Þ
0
0
0:38607ðz 1 0:4182Þðz 2 0:2479 3 1022 Þ
8.4 Poles and zeros of multivariable systems
For this square 2-input-2-output system, the determinant of the transfer function is
ðz 1 0:1067Þðz 2 0:3012Þðz 1 0:4182Þðz 2 0:2479 3 1022 Þðz20:2466Þ2
ðz20:2466Þ4 ðz20:2479 3 1022 Þ2 ðz20:3012Þ2
5
ðz 1 0:1067Þðz 1 0:4182Þ
ðz20:2466Þ2 ðz 2 0:2479 3 1022 Þðz 2 0:3012Þ
The roots of the numerator are the zeros { 2 0.1067, 20.4182}.
The same answer is obtained using the MATLAB command tzero.
2
Rober, S.J., Shin, Y.C., 1995. Modeling and control of CNC machines using a PC-based open architecture
controller, Mechatronics 5 (4), 401420.
EXAMPLE 8.16
Determine the zeros of the transfer function matrix
3
1
6
ðz 2 1Þðz 2 0:5Þ 7
7
6
7
GðzÞ 5 6
7
6
z
2
0:1
1
5
4
ðz 2 0:2Þðz 2 0:5Þ
z 2 0:2
2
1
z21
Solution
From Example 8.14, we know that the poles of the system are at {0.2, 0.5, 0.5, 1}. We
rewrite the transfer function matrix in the form
"
GðzÞ 5
ðz 2 0:2Þðz20:5Þ2
ðz 2 0:2Þðz 2 0:5Þ
#
ðz20:5Þ2 ðz 2 1Þ
ðz 2 0:1Þðz 2 0:5Þðz 2 1Þ
ðz 2 0:2Þðz20:5Þ ðz 2 1Þ
2
Zeros are the roots of the greatest common divisor of all minors of order equal to 2,
which in this case is simply the determinant of the transfer function matrix. The determinant of the matrix with cancellation to reduce the denominator to the characteristic polynomial is
det½GðzÞ 5
ðz 2 0:5Þ2 2ðz 2 0:1Þ
z2 2 2z 1 0:35
5
2
ðz 2 0:2Þðz 2 0:5Þ ðz 2 1Þ ðz 2 0:2Þðz 2 0:5Þ2 ðz 2 1Þ
The roots of the numerator yield zeros at {1.8062, 20.1938}.
The same answer is obtained using the MATLAB command zero.
8.4.2 Zeros from statespace models
System zeros can be obtained from a statespace realization using the definition
of the zero. However, this is complicated by the possibility of pole-zero
323
324
CHAPTER 8 Properties of StateSpace Models
cancellation if the system is not minimal. We rewrite the zero condition (8.25) in
terms of the statespace matrices as
Gðz0 Þw 5 C½z0 In 2A21 Bw 1 Dw
5 Cxw 1 Dw 5 0
(8.27)
21
xw 5 ½z0 In 2A Bw
where xw is the state response for the system with input w. We can now rewrite
(8.27) in the following form, which is more useful for numerical solution:
2ðz0 In 2 AÞ B
C
D
0
xw
5
w
0
(8.28)
The matrix in (8.28) is known as Rosenbrock’s system matrix. Equation (8.28)
can be solved for the unknowns provided that a solution exists.
Note that the zeros are invariant under similarity transformation because
2ðz0 In 2 Tr21 ATr Þ Tr21 B
D
CTr
0
2ðz0 In 2 AÞ B Tr xw
xw
Tr21 0
5
5
w
w
0
C
D
0 Il
In Chapter 9, we discuss state feedback where the control input is obtained as
a linear combination of the measured state variables. We show that the zeros
obtained from (8.28) are also invariant under state feedback. Hence, the zeros are
known as invariant zeros.
The invariant zeros include decoupling zeros if the system is not minimal.
If these zeros are removed from the set of invariant zeros, then the remaining
zeros are known as transmission zeros. In addition, not all decoupling zeros can
be obtained from (8.28). We therefore have the following relation:
fsystem zerosg 5 ftransmission zerosg 1 finputdecoupling zerosg
1 foutputdecoupling zerosgfinputoutputdecoupling zerosg
Note that some authors refer to invariant zeros as transmission zeros and do not
use the term “invariant zeros.”
MATLAB calculates zeros from Rosenbrock’s matrix of a statespace model
p using the command
.. zeroðpÞ
The MATLAB manual identifies the result as transmission zeros, but in our terminology this refers to invariant zeros. Although the command accepts a transfer
function matrix, its results are based on a realization that need not be minimal
and may include superfluous zeros that would not be obtained using the procedure
described in Section 8.4.1.
8.5 Statespace realizations
EXAMPLE 8.17
Consider the system
2
3 2
32 x ðkÞ 3 2 1
x1 ðk 1 1Þ
1
20:4 0
0
6
7
6
7 6
3
0 54 x2 ðkÞ 5 1 4 0
4 x2 ðk 1 1Þ 5 5 4 0
0
0
22
1
x3 ðk 1 1Þ
x3 ðkÞ
2
3
x1 ðkÞ
6
7
yðkÞ 5 ½ 1 1 0 4 x2 ðkÞ 5
x3 ðkÞ
0
3
7
0 5uðkÞ
1
32
3 2 3
2
xw1
0
20:4 2 z0
0
0
1
0
3
76 x 7 6 0 7
6
w2
7
B " # 6
2ðz0 In 2 AÞ
6
7
7
6
0
3 2 z0
0
0
0 76
7 6 7
6
6
7 xw
76 xw3 7 5 6 0 7
56
4 5
7 6 7
6
6
0
0
22 2 z0
1
1 7
w
76
7 6 7
6
C
D
4 54 w1 5 4 0 5
0
0
22
0
0
0
w2
2
From the second row, we have the invariant zero z0 5 3, since this value makes the
matrix rank deficient. This is also an input-decoupling zero because it corresponds to an
uncontrollable mode. The system has an output-decoupling zero at 22, but this cannot be
determined from Rosenbrock’s system matrix.
8.5 Statespace realizations
Statespace realizations can be obtained from input-output time domain or
z-domain models. Because every system has infinitely many statespace realizations, we only cover a few standard canonical forms that have special desirable
properties. These realizations play an important role in digital filter design or controller implementation. Because difference equations are easily transformable to
z-transfer functions and vice versa, we avoid duplication by obtaining realizations
from either the z-transfer functions or the difference equations.
In this section, we mainly discuss SISO transfer functions of the form
cn zn 1 cn21 zn21 1 . . . 1 c1 z 1 c0
zn 1 an21 zn21 1 a1 z 1 a0
cn21 zn21 1 cn22 zn22 1 . . . 1 c1 z 1 c0
5 cn 1
zn 1 an21 zn21 1 . . . 1 a1 z 1 a0
GðzÞ 5
(8.29)
where the leading numerator coefficients cn and cn21 can be zero and cn 5 cn, or
the corresponding difference equation
yðk 1 nÞ 1 an21 yðk 1 n 2 1Þ 1 . . . 1 a1 yðk 1 1Þ 1 a0 yðkÞ
5 cn uðk 1 nÞ 1 cn21 uðk 1 n 2 1Þ 1 . . . 1 c1 uðk 1 1Þ 1 c0 uðkÞ
(8.30)
325
326
CHAPTER 8 Properties of StateSpace Models
8.5.1 Controllable canonical realization
The controllable canonical realization is so called because it possesses the
property of controllability. The controllable form is also known as the phase
variable form or as the first controllable form. A second controllable form,
known as controller form, is identical in structure to the first but with the state
variables numbered backward. We begin by examining the special case of a
difference equation whose RHS is the forcing function at time k. This corresponds to a transfer function with unity numerator. We then use our results to
obtain realizations for a general SISO linear discrete-time system described by
(8.29) or (8.30).
Systems with no input differencing
Consider the special case of a system whose difference equation includes the
input at time k only. The difference equation considered is of the form
yðk 1 nÞ 1 an21 yðk 1 n 2 1Þ 1 . . . 1 a1 yðk 1 1Þ 1 a0 yðkÞ 5 uðkÞ
(8.31)
Define the state vector
xðkÞ 5 x1 ðkÞ x2 ðkÞ
. . . xn21 ðkÞ xn ðkÞ
T
T
(8.32)
xn ðk 1 1Þ 5 2an21 xn ðkÞ 2 . . . 2 a1 x2 ðkÞ 2 a0 x1 ðkÞ 1 uðkÞ
(8.33)
5 yðkÞ
yðk11Þ
. . . yðk1n22Þ yðk1n21Þ
Hence, the difference equation (8.31) can be rewritten as
Using the definitions of the state variables and equation (8.33), we obtain the
matrix state equation
3 2
x1 ðk 1 1Þ
6 x2 ðk 1 1Þ 7 6
6
7 6
6
756
U
6
7 6
4 xn21 ðk 1 1Þ 5 4
xn ðk 1 1Þ
2
0
0
...
0
2a0
1
0
...
0
2a1
...
...
...
...
...
0
0
...
0
2an22
32
3 2 3
x1 ðkÞ
0
0
6 x2 ðkÞ 7 6 0 7
0 7
76
7 6 7
6
7 6 7
... 7
76 U 7 1 6 U 7uðkÞ
5
4
xn21 ðkÞ 5 4 0 5
1
2an21
xn ðkÞ
1
(8.34)
and the output equation
2
yðkÞ 5 ½ 1
0
...
0
3
x1 ðkÞ
6 x2 ðkÞ 7
6
7
7
0 6
6 U 7
4 xn21 ðkÞ 5
xn ðkÞ
(8.35)
8.5 Statespace realizations
The statespace equations can be written more concisely as
2
3
2
3
0n21 3 1
In21
0n21 3 1
xðk 1 1Þ 5 4 5xðkÞ 1 4 5uðkÞ
1
2a1 ? ? 2an21
2a0
yðkÞ 5 1
013n21 xðkÞ
(8.36)
Clearly, the statespace equations can be written by inspection from the transfer function
GðzÞ 5
1
zn 1 an21 zn21 1 . . . 1 a1 z 1 a0
(8.37)
or from the corresponding difference equation because either includes all the
needed coefficients, ai, i 5 0, 1, 2, . . . , n 2 1.
EXAMPLE 8.18
Obtain the controllable canonical realization of the difference equation
yðk 1 3Þ 1 0:5yðk 1 2Þ 1 0:4yðk 1 1Þ 2 0:8yðkÞ 5 uðkÞ
using basic principles; then show how the realization can be written by inspection from
the transfer function or the difference equation.
Solution
Select the state vector
xðkÞ 5 x1 ðkÞ x2 ðkÞ x3 ðkÞ
T
5 yðkÞ yðk11Þ yðk12Þ
T
and rewrite the difference equation as
x3 ðk 1 1Þ 5 20:5x3 ðkÞ 2 0:4x2 ðkÞ 1 0:8x1 ðkÞ 1 uðkÞ
Using the definitions of the state variables and the difference equation, we obtain the
statespace equations
32
3 2 3
0
1
0
x1 ðkÞ
0
7
6
0
0
1 76
6
7
7 6 7
1
0 5uðkÞ
ðkÞ
x
4 x2 ðk 1 1Þ 5 5 6
4
5
4
2
4 5
ðkÞ
1
x3 ðk 1 1Þ
x
3
0:8 20:4 20:5
2
x1 ðk 1 1Þ
3
2
2
yðkÞ5 1
0
3
x1 ðkÞ
6
7
0 4 x2 ðkÞ 5
x3 ðkÞ
because the system is of order n 5 3, n 2 1 5 2. Hence, the upper half of the state matrix
includes a 2 3 1 zero vector next to a 2 3 2 identity matrix. The input matrix is a column
vector because the system is SI, and it includes a 2 3 1 zero vector and a unity last entry.
The output matrix is a row vector because the system is single output (SO), and it includes
a 1 3 2 zero vector and a unity first entry. With the exception of the last row of the state
327
328
CHAPTER 8 Properties of StateSpace Models
matrix, all matrices in the statespace description are completely determined by the order
of the system. The last row of the state matrix has entries equal to the coefficients of the
output terms in the difference equation with their signs reversed. The same coefficients
appear in the denominator of the transfer function
GðzÞ 5
1
z3 1 0:5z2 1 0:4z 2 0:8
Therefore, the statespace equations for the system can be written by inspection from the
transfer function or input-output difference equation.
Systems with input differencing
We now use the results from the preceding section to obtain the controllable
canonical realization of the transfer function of equation (8.29). We assume that
the constant term has been extracted from the transfer function, if necessary, and
that we are left with the form
YðzÞ
5 cn 1 Gd ðzÞ
UðzÞ
(8.38)
cn21 zn21 1 cn22 zn22 1 . . . 1 c1 z 1 c0
zn 1 an21 zn21 1 . . . 1 a1 z 1 a0
(8.39)
GðzÞ 5
where
Gd ðzÞ 5
We next consider a transfer function with the same numerator as Gd but with
unity numerator, and we define a new variable p(k) whose z-transform satisfies
PðzÞ
1
5 n
n21
UðzÞ z 1 an21 z 1 . . . 1 a1 z 1 a0
(8.40)
The state equation of a system with the preceding transfer function can be written
by inspection, with the state variables chosen as
xðkÞ 5 x1 ðkÞ
5 pðkÞ
x2 ðkÞ
pðk11Þ
...
xn21 ðkÞ
xn ðkÞ
. . . pðk1n22Þ
T
pðk1n21Þ
T
(8.41)
This choice of state variables is valid because none of the variables can be written
as a linear combination of the others. However, we have used neither the numerator
of the transfer function nor the constant cn. Nor have we related the state variables
to the output. So we multiply (8.39) by U(z) and use (8.40) to obtain
YðzÞ 5 cn UðzÞ 1 Gd ðzÞUðzÞ
5 cn UðzÞ 1
cn21 zn21 1 cn22 zn22 1 . . . 1 c1 z 1 c0
UðzÞ
zn 1 an21 zn21 1 . . . 1 a1 z 1 a0
5 cn UðzÞ 1 cn21 zn21 1 cn22 zn22 1 . . . 1 c1 z 1 c0 PðzÞ
(8.42)
8.5 Statespace realizations
Then we inverse z-transform and use the definition of the state variables to obtain
yðkÞ 5 cn uðkÞ 1 cn21 pðk 1 n 2 1Þ 1 cn22 pðk 1 n 2 2Þ 1 . . . 1 c1 pðk 1 1Þ 1 c0 pðkÞ
5 cn uðkÞ 1 cn21 xn ðkÞ 1 cn22 xn21 ðkÞ 1 . . . 1 c1 x2 ðkÞ 1 c0 x1 ðkÞ
(8.43)
Finally, we write the output equation in the matrix form
2
3
x1 ðkÞ
6 x2 ðkÞ 7
6
7
6 U 7
yðkÞ 5 ½ c0 c1 . . . cn22 cn21 6
7 1 duðkÞ
6 U 7
4 x ðkÞ 5
n21
xn ðkÞ
(8.44)
where d 5 cn.
As in the preceding section, the statespace equations can be written by
inspection from the difference equation or the transfer function.
A simulation diagram for the system is shown in Figure 8.4. The simulation
diagram shows how the system can be implemented in terms of summer, delay,
and scaling operations. The number of delay elements needed for implementation
is equal to the order of the system. In addition, two summers and at most 2n 1 1
gains are needed. These operations can be easily implemented using a microprocessor or digital signal processing chip.
cn
c2
x2(k)
u(k)
+
T
xn(k)
T
T
T
−an−1
−an−2
−a0
FIGURE 8.4
Simulation diagram for the controllable canonical realization.
T
c1
x1(k)
y(k)
c0
+
329
330
CHAPTER 8 Properties of StateSpace Models
EXAMPLE 8.19
Write the statespace equations in controllable canonical form for the following transfer
functions:
1. GðzÞ 5
0:5ðz 2 0:1Þ
z3 1 0:5z2 1 0:4z 2 0:8
2. GðzÞ 5
z4 1 0:1z3 1 0:7z2 1 0:2z
z4 1 0:5z2 1 0:4z 2 0:8
Solution
1. The transfer function has the same denominator as that shown in Example 8.18.
Hence, it has the same state equation. The numerator of the transfer function can be
expanded as (20.05 1 0.5z 1 0z2), and the output equation is of the form
2
3
x1 ðkÞ
4
yðkÞ 5 ½20:05 0:5 0 x2 ðkÞ 5
x3 ðkÞ
2. The transfer function has the highest power of z equal to 4, in both the numerator
and denominator. We therefore begin by extracting a constant equal to the numerator z4
coefficient to obtain
GðzÞ 5 1 1
511
0:1z3 1ð0:7 2 0:5Þz2 1ð0:2 2 0:4Þz 2 ð20:8Þ
z4 1 0:5z2 1 0:4z 2 0:8
0:1z3 1 0:2z2 2 0:2z 1 0:8
z4 1 0:5z2 1 0:4z 2 0:8
Now we can write the statespace equations by inspection as
32
3 2
3 2 3
2
x1 ðkÞ
0
1
0
0
0
x1 ðk 1 1Þ
7
6 7
6 x2 ðk 1 1Þ 7 6 0
6
0
1
0 76 x2 ðkÞ 7
756
7 1 6 0 7uðkÞ
6
4 x3 ðk 1 1Þ 5 4 0
0
0
1 54 x3 ðkÞ 5 4 0 5
0:8 20:4 20:5 0
1
x4 ðk 1 1Þ
x4 ðkÞ
3
2
x1 ðkÞ
6 x2 ðkÞ 7
7
yðkÞ 5 ½ 0:8 20:2 0:2 0:1 6
4 x3 ðkÞ 5 1 uðkÞ
x4 ðkÞ
Theorem 8.5 can be used to show that any system in controllable form is
actually controllable. The proof is straightforward and is left as an exercise (see
Problem 8.15).
8.5.2 Controllable form in MATLAB
MATLAB gives a canonical form that is almost identical to the controllable form
of this section with the command
.. ½A;B;C;D 5 tf2ssðnum; denÞ
8.5 Statespace realizations
Using the command with the system described in Example 8.19(2) gives the
statespace equations
2
3 2
x1 ðk 1 1Þ
0
6 x2 ðk 1 1Þ 7 6 1
6
7 6
4 x3 ðk 1 1Þ 554 0
x4 ðk 1 1Þ
0
20:5
0
1
0
yðkÞ 5 ½ 0:1
0:2
3 2 3
32
20:4 0:8
1
x1 ðkÞ
6 x2 ðkÞ 7 6 0 7
0
0 7
7 1 6 7uðkÞ
76
0
0 54 x3 ðkÞ 5 4 0 5
x4 ðkÞ
1
0
0
2
3
x1 ðkÞ
6 x2 ðkÞ 7
7
20:2 0:8 6
4 x3 ðkÞ 5 1 uðkÞ
x4 ðk Þ
The equations are said to be in controller form. By drawing a simulation
diagram for this system and then for the system described in Example 8.19 (2),
we can verify that the two systems are identical. The state variables for the form
in MATLAB are simply numbered in the reverse of the order used in this text
(see Figure 8.4); that is, the variables xi in the MATLAB model are none other
than the variables xn2i11, i 5 1, 2, . . . , n.
8.5.3 Parallel realization
Parallel realization of a transfer function is based on the partial fraction expansion
Gd ðzÞ 5 d 1
n
X
cn21 zn21 1 cn22 zn22 1 . . . 1 c1 z 1 c0
Ki
5
d
1
n
n21
.
.
.
z 1 an21 z 1 1 a1 z 1 a0
z 1 pi
i51
(8.45)
The expansion is represented by the simulation diagram of Figure 8.5. The
summation in (8.45) gives the parallel configuration shown in the figure, which
justifies the name parallel realization.
A simulation diagram can be obtained for the parallel realization by
observing that the z-transfer functions in each of the parallel branches can be
rewritten as
1
z21
5
1 2 ð2pi z21 Þ
z 1 pi
which can be represented by a positive feedback loop with forward transfer
function z21 and feedback gain 2pi, as shown in Figure 8.6. Recall that z21 is
simply a time delay so that a physical realization of the transfer function in terms
of constant gains and fixed time delays is now possible.
331
CHAPTER 8 Properties of StateSpace Models
X1(z)
1
z + p1
K1
…
…
Y(z)
U(z)
Xn(z)
1
z + pn
+
Kn
d
FIGURE 8.5
Block diagram for parallel realization.
X1(z)
z–1
+
K1
−p1
Y(z)
…
U(z)
…
332
Xn(z)
z –1
+
+
Kn
−pn
d
FIGURE 8.6
Simulation diagram for parallel realization.
We define the state variables as the outputs of the first-order blocks and
inverse z-transform to obtain the state equations
xi ðk 1 1Þ 5 2pi xi ðkÞ 1 uðkÞ;
i 5 1; 2; . . . ; n
(8.46)
The output is given by the summation
yðkÞ 5
n
X
i51
Ki xi ðkÞ 1 duðkÞ
(8.47)
8.5 Statespace realizations
Equations (8.46) and (8.47) are equivalent to the statespace representation
2
x1 ðk 1 1Þ
3 2
6
7 6
6 x2 ðk 1 1Þ 7 6
6
7 6
6
75 6
U
6
7 6
6
7 6
4 xn21 ðk 1 1Þ 5 4
xn ðk 1 1Þ
2p1
0
^
0
0
...
0
2p2 . . .
^ ...
0
^
0
2 3
1
76
7 6 7
0 76 x2 ðkÞ 7 6 1 7
76
7 6 7
6
7 6 7
^ 7
76 ^ 7 1 6 ^ 7uðkÞ
76
7 6 7
0 54 xn21 ðkÞ 5 4 1 5
1
2pn
xn ðkÞ
. . . 2pn21
...
0
0
0
32
0
2
x1 ðkÞ
3
x1 ðkÞ
6
7
6 x2 ðkÞ 7
6
7
7
K n 6
6 ^ 7 1 duðkÞ
6
7
4 xn21 ðkÞ 5
xn ðkÞ
yðkÞ 5 ½ K1 K2 . . . Kn21
EXAMPLE 8.20
Obtain a parallel realization for the transfer function
GðzÞ 5
2z2 1 2z 1 1
z2 1 5z 1 6
Solution
We first write the transfer function in the form
GðzÞ 5 2 1
522
2z2 1 2z 1 1 2 2ðz2 1 5z 1 6Þ
z2 1 5z 1 6
8z 1 11
ðz 1 2Þðz 1 3Þ
Then we obtain the partial fraction expansion
GðzÞ 5 2 1
5
13
2
z12 z13
Finally, we have the statespace equations
"
x1 ðk 1 1Þ
x2 ðk 1 1Þ
#
"
5
#"
22
0
0
23
"
yðkÞ 5 ½ 5
3
213 " #
1
1
uðkÞ
1
x2 ðkÞ
x1 ðkÞ
x1 ðkÞ
x2 ðkÞ
#
#
1 2uðkÞ
(8.48)
333
334
CHAPTER 8 Properties of StateSpace Models
X1(z)
1
z+2
5
Y(z)
+
U(z)
X2(z)
1
z +3
−13
2
FIGURE 8.7
Block diagram for Example 8.20.
z−1
+
X1(z)
5
−2
Y(z)
+
U(z)
z−1
+
X2(z)
−13
−3
2
FIGURE 8.8
Simulation diagram for Example 8.20.
The block diagram for the parallel realization is shown in Figure 8.7, and the simulation diagram is shown in Figure 8.8. Clearly, the system is unstable with two eigenvalues
outside the unit circle.
Parallel realization for MIMO systems
For MIMO systems, a parallel realization can be obtained by partial fraction
expansion as in the SISO case. The method requires that the minimal polynomial
of the system have no repeated roots. The partial fractions in this case are constant matrices.
The partial fraction expansion is in the form
Gd ðzÞ 5 D 1
n
X
Pn21 zn21 1 Pn22 zn22 1 . . . 1 P1 z 1 P0
Ki
5
D
1
n
n21
z 1 an21 z 1 . . . 1 a1 z 1 a0
z
1
pi
i51
(8.49)
8.5 Statespace realizations
where D, Pi, and Ki, i 5 1, . . . , n 2 1 are l 3 m matrices. Each of the matrices Ki
must be decomposed into the product of two full-rank matrices: an input component matrix Bi and an output component matrix Ci. We write the partial fraction
matrices in the form
Ki 5 Ci Bi ;
i 5 1; . . . ; n
(8.50)
For full-rank component matrices, their order is dictated by the rank of the
matrix Ki. This follows from the fact that the rank of the product is at most equal
to the minimum dimension of the components. For rank (Ki) 5 ri, we have Ci as
an l 3 ri matrix and Bi as an ri 3 m matrix. The rank of the matrix Ki represents
the minimum number of poles pi needed for the parallel realization. In fact, it can
be shown that this realization is indeed minimal. The parallel realization is given
by the quadruple
2
3
2
3
B1
0
...
0
0
2Ir1 p1
6 0
6 B 7
2Ir2 p2 . . .
0
0 7
6
7
6 2 7
6
6
7
7
7 B56 ^ 7
^
^
.
.
.
^
^
A56
6
7
6
7
(8.51)
6
7
6
7
4 0
4 Bn21 5
0
. . . 2Irn21 pn21
0 5
0
C 5 C1
...
0
C2
...
Cn21
2Irn pn
0
Cn
Bn
D
EXAMPLE 8.21
Obtain a parallel realization for the transfer function matrix of Example 8.14:
3
1
ðz 2 1Þðz 2 0:5Þ 7
7
7
7
z 2 0:1
1
5
ðz 2 0:2Þðz 2 0:5Þ
z 2 0:2
2
1
z21
6
6
GðzÞ 5 6
6
4
Solution
The minimal polynomial (z 2 0.2)(z 2 0.5)(z 2 1) has no repeated roots. The partial fraction
expansion of the matrix is
2
6
6
GðzÞ 5 6
6
4
1
z21
z 2 0:1
ðz 2 0:2Þðz 2 0:5Þ
2
0
6 1
42
3
5
0
3
2
0
64
4
3
7
15
z 2 0:2
1
3
1
ðz 2 1Þðz 2 0:5Þ 7
7
7
7
1
5
z 2 0:2
22
3
7
0 5
z 2 0:5
1
1
2
0 0
z21
335
CHAPTER 8 Properties of StateSpace Models
The ranks of the partial fraction coefficient matrices given with the corresponding poles
are (1, 0.2), (2, 0.5), and (1, 1). The matrices can be factorized as
3
2
" #"
#
1
7
6
0
2
1
75
6 1
3
42
15
1
3
3
3
2
2
"
# 0 22
0 22
7
7
6
1 0 6
7
7
64
64
4
0 55 0 1 4
0 5
3
3
"
# " #
1
1 2
1 2
5
0
0 0
0
0
The parallel realization is given by the quadruple
2
2
0:2
6 0
6
A56
4 0
0
C5
0
1
0
0
0
0:5
0
0
0:5
0
0
1 0
0 1
1
0
3
07
7
7
05
1
2
6
6
6
6 0
B56
6 4
6
6
4 3
1
3
1
3
1
7
7
7
22 7
7
7
7
0 7
5
2
D 5 0232
Note that the realization is minimal as it is fourth order, and in Example 8.14, the system
was determined to have four poles.
8.5.4 Observable form
The observable realization of a transfer function can be obtained from the
controllable realization using the following steps:
1. Transpose the state matrix A.
2. Transpose and interchange the input matrix B and the output matrix C.
Clearly, the coefficients needed to write the statespace equations all appear in
the transfer function, and all equations can be written in observable form directly.
Using (8.36) and (8.44), we obtain the realization
2
2
3
3
013 n21 2a0
c0
6
6 c 7
2a1 7
6
6 1 7
7
xðk 1 1Þ 5 6
7xðkÞ 1 6
7uðkÞ
4 In21
4 ^ 5
5
^
(8.52)
2an21
cn21
yðkÞ5 013n21
336
1 xðkÞ 1 duðkÞ
8.5 Statespace realizations
u(k)
c0
+
T
x1(k)
−a0
c1
cn−2
+
+
−a1
−an−2
cn−1
T
xn−1(k)
+
d
T
xn(k)
+
y(k)
−an−1
FIGURE 8.9
Simulation diagram for the observable canonical realization.
The simulation diagram for the observable realization is shown in Figure 8.9.
Note that it is possible to renumber the state variables in the simulation diagram to
obtain an equivalent realization known as observer form. However, the resulting
realization will clearly have different matrices from those of (8.52). The second
observable realization can be obtained from the controller form by transposing
matrices as in the two preceding steps.
The two observable realizations can also be obtained from the basic principles;
however, the derivations have been omitted in this text and are left as an exercise.
Theorem 8.5 can be used to show that any system in observable form is actually
observable. The proof is straightforward and is left as an exercise (see Problem 8.16).
EXAMPLE 8.22
Write the statespace equations in observable canonical form for the transfer function of
Example 8.19 (2).
GðzÞ 5
z4 1 0:1z3 1 0:7z2 1 0:2z
z4 1 0:5z2 1 0:4z 2 0:8
Solution
The transfer function can be written as a constant plus a transfer function with numerator
order less than the denominator order as in Example 8.19 (2). Then, using (8.52), we
obtain the following:
32
3
3 2
3 2
2
x1 ðkÞ
0 0 0
0:8
0:8
x1 ðk 1 1Þ
6 x2 ðk 1 1Þ 7 6 1 0 0 20:4 76 x2 ðkÞ 7 6 20:2 7
76
7
7 6
7 6
6
4 x3 ðk 1 1Þ 5 5 4 0 1 0 20:5 54 x3 ðkÞ 5 1 4 0:2 5uðkÞ
0 0 1
0
0:1
x4 ðk 1 1Þ
x4 ðkÞ
3
2
x1 ðkÞ
6 x2 ðkÞ 7
7 1 uðkÞ
6
yðkÞ 5 ½ 0 0 0 1 4
x3 ðkÞ 5
x4 ðkÞ
337
CHAPTER 8 Properties of StateSpace Models
The same realization is obtained from the controllable realization of Example 8.19 (2) by
transposing the state matrix and transposing, and then interchanging the matrices B and C.
8.6 Duality
The concepts of controllability (stabilizability) and observability (detectability)
are often referred to as duals. This term is justified by Theorem 8.12.
THEOREM 8.12
The pair (A, B) is controllable (stabilizable) if and only if (AT, BT) is observable (detectable).
The system (A, C) is observable (detectable) if and only if (AT, CT) is controllable (stabilizable).
PROOF
The relevant controllability and observability matrices are related by the equations
3T
2
BT
6 7
7
6
7
6 BT AT
7
6
7 5 O T ðAT ; BT Þ
CðA; BÞ 5 B AB . . . An21 B 5 6
7
6
7
6
^
7
6
4 5
T
T n21
B ðA Þ
3T
B
6 7
7
6
6 BA 7
7
6
7 5 O T ðA; BÞ
56
7
6
6 ^ 7
7
6
4 5
n21
BA
AT BT
CðAT ; BT Þ 5 BT
2
338
...
ðAT Þn21 BT
The proof follows from the equality of the rank of any matrix and the rank of its transpose. The statements regarding detectability and stabilizability are true because (i) a matrix
and its transpose have the same eigenvalues and (ii) the right eigenvectors of the transpose
of the matrix when transposed give the left eigenvectors of the original matrix. Details of the
proof are left as an exercise (Problem 8.28).
EXAMPLE 8.23
Show that the reducible transfer function
GðzÞ 5
0:3ðz 2 0:5Þ
ðz 2 1Þðz 2 0:5Þ
has a controllable but unobservable realization and an observable but uncontrollable
realization.
8.7 Hankel realization
Solution
The transfer function can be written as
0:3z 2 0:15
z2 2 1:5z 1 0:5
GðzÞ 5
The controllable realization for this system is
x1 ðk 1 1Þ
5
x2 ðk 1 1Þ
x1 ðkÞ
0
1
uðkÞ
x2 ðkÞ
1
x ðkÞ
y 5 ½20:15 0:3 1
x2 ðkÞ
0
1
20:5 1:5
The observability matrix for this realization is
2
3 2
3
C
20:15 0:3
4
5
4
5
O5
5
CAd
20:15 0:3
The observability matrix has rank 1. The rank deficit is 2 2 1 5 1, corresponding to one
unobservable mode.
Transposing the state, output, and input matrices and interchanging the input and
output matrices gives the observable realization
0
x1 ðk 1 1Þ
5
x2 ðk 1 1Þ
1
20:5
1:5
yðkÞ 5 ½ 0 1 x1 ðkÞ
x2 ðkÞ
x1 ðkÞ
1
x2 ðkÞ
20:15
uðkÞ
0:3
By duality, the realization is observable but has one uncontrollable mode.
8.7 Hankel realization
In this section we obtain a realization for a MIMO system using singular value
decomposition (see Appendix III). Consider a transfer function matrix expanded
in the form
GðzÞ 5
N
X
GðkÞz2k
(8.53)
k50
The terms of the infinite expansion are known as the Markov parameter matrices
of the system. The terms can be obtained using
Gð0Þ 5 lim GðzÞ
z-N
GðiÞ 5 lim zi fGðzÞ 2
z-N
i21
X
k50
GðkÞz2k g; i 5 1; 2; . . . ; N
(8.54)
339
340
CHAPTER 8 Properties of StateSpace Models
From (7.95) we have the direct transmission matrix
D 5 Gð0Þ
(8.55)
and the remaining terms of the expansion must be used to obtain a realization for
the system. To this end, we first define the Hankel matrix as
2
3
?
GðpÞ
5
&
^
? Gð2p 2 1Þ
Gð1Þ
Hðp; pÞ 5 4 ^
GðpÞ
(8.56)
We also need the shifted Hankel matrix
2
Gð2Þ
?
^
&
Hs ðp; pÞ 5 4
Gðp 1 1Þ ?
3
Gðp 1 1Þ
5
^
Gð2pÞ
(8.57)
Let p be the degree of the least common denominator of the transfer function
matrix. Then we know that the order of the minimal realization is n # p . For the
n 3 n block Hankel matrix, we see from the impulse response expansion (7.95) that
2
3
C
6 CA 7
7
Hðn; nÞ 5 6
4 ^ 5 B
CAn21
AB
?
An21 B 5 OC
(8.58)
2
3
C
6 CA 7
7
Hs ðn; nÞ 5 6
4 ^ 5A B
CAn21
AB
?
An21 B 5 O A C
(8.59)
So that the Hankel matrix of the appropriate size is the product of the observability matrix and the controllability matrix of the minimal realization. The problem is that we do not know the order of this realization in advance. However,
with the condition n # p, we can obtain the singular value decomposition of the
Hankel matrix
Hðp; pÞ 5 U
Σ
0n3ðp2nÞ
0ðp2nÞ3n
0ðp2nÞ3ðp2nÞ
V
(8.60)
where U and V are orthogonal matrices and Σ 5 diag{σ1, . . . , σn} is a diagonal
matrix with positive entries. The n 3 n block Hankel matrix is the submatrix given by
Hðn; nÞ 5 U1:n Σ 1=2 Σ1=2 V1:n
(8.61)
8.7 Hankel realization
n
o
1=2
1=2
With U1:n denoting the first n columns of the matrix and Σ1=2 5 diag σ1 ?σn ,
we can now select the matrices
C 5 Σ1=2 V1:n
(8.62)
O 5 U1:n Σ1=2
(8.63)
B 5 first m columns of C
(8.64)
C 5 first l rows of O
(8.65)
A 5 O 1 Hs ðn; nÞ C 1
(8.66)
From the above matrices we have
where C 1 denotes the pseudoinverse of the matrix.
We now summarize the procedure to obtain the Hankel realization.
PROCEDURE 8.1
1. Find the order p of the least common denominator of all the entries of the transfer
function matrix.
2. Obtain the first 2p Markov parameter matrices GðkÞ; k 5 0; 1; ?; 2p and form the p 3 p
block Hankel matrix H(p, p).
3. Obtain the singular value decomposition of the matrix H(p, p) and determine the
order of the minimal realization n.
4. Calculate the controllability and observability matrices of the minimal realization
using the first n right and left singular vectors and the nonzero singular values.
5. Obtain the n 3 n block Hankel matrix Hs(n, n).
6. Obtain a minimal realization using (8.55) and (8.64) to (8.66).
EXAMPLE 8.24
Obtain the Hankel realization of the system of Example 8.14.
Solution
The transfer function matrix
3
1
6
ðz 2 1Þðz 2 0:5Þ 7
7
6
7
GðzÞ 5 6
7
6
z 2 0:1
1
5
4
ðz 2 0:2Þðz 2 0:5Þ
z 2 0:2
2
1
z21
has the least common denominator
ðz 2 0:2Þðz 2 0:5Þðz 2 1Þ
341
342
CHAPTER 8 Properties of StateSpace Models
and thus p 5 3 and 2p 5 6. The first six Markov parameter matrices are
Gð0Þ 5 lim GðzÞ 5 0232
z-N
Gð1Þ 5 lim zfGðzÞ 2 Gð0Þg 5
z-N
1
1
0
1
Gð2Þ 5 lim z2 fGðzÞ 2 z21 Gð1Þ 2 Gð0Þg 5
z-N
1
1
0:6 0:2
Gð3Þ 5 lim z3 fGðzÞ 2 z22 Gð2Þ 2 z21 Gð1Þ 2 Gð0Þg 5
z-N
1
1:5
0:32 0:04
Gð4Þ 5 lim z4 fGðzÞ 2 z23 Gð3Þ 2 z22 Gð2Þ 2 z21 Gð1Þ 2 Gð0Þg 5
z-N
1
1:75
0:164 0:008
Gð5Þ 5 lim z5 fGðzÞ 2 z24 Gð4Þ 2 z23 Gð3Þ 2 z22 Gð2Þ 2 z21 Gð1Þ 2 Gð0Þg 5
z-N
We form the Hankel matrix using the Markov parameters
2
1 0
1
1
2
3 6
1 1
0:6 0:2
6
Gð1Þ Gð2Þ Gð3Þ
6
6
1
1
1:5
Hð3; 3Þ 5 4 Gð2Þ Gð3Þ Gð4Þ 5 5 6 1
6 0:6 0:2
0:32 0:04
Gð3Þ Gð4Þ Gð5Þ
6
4 1
1:5
1
1:75
0:32 0:04 0:164 0:008
1
1:875
0:0828 0:0016
1
1:5
0:32 0:04
3
7
7
7
1
1:75 7
7
0:164 0:008 7
7
1
1:875 5
0:0828 0:0016
Then we determine its rank and singular value decomposition using the MATLAB command
n 5 rank(Hank); % Find the rank of the Hankel matrix
[L,Sig,R] 5 svd(Hank);% Singular value decomposition Hank 5 L Sig R';
ans 5
4
Using the singular value decomposition of the Hankel matrix, we obtain the controllability matrix
C 5 Σ1=2 V1:n
2
20:7719
6 20:4768
6
56
4 0:7110
20:1441
20:8644
20:7463
21:1275
20:7294
21:2712
20:7202
20:9415
20:1550
0:0267
0:0588
0:3521
0:1763
20:4946
0:4158
20:3515
0:2322
20:1721
0:1332
0:0909
0:0867
20:1812
0:1150
20:0560
0:1118
21:3455
0:4831 7
7
7
20:0608 5
0:0588
and the observability matrix
2
6
6
6
6
6
6
O 5 U1:n Σ1=2 5 6
6
6
6
6
4
21:
09260:
67400:0
42060:1
569 2 1:
04800:0 064 2 0:1
21:4680 20:1085
20:0661 20:1323
21:5306 20:2358
20:0318 20:0603
3
3
65580:
689 2 0:3560 2 0:94110: 7
7
7
34290:
14630:
7
453 2 0:
32350:4292 2 0:1969 7
7
7
20:2651
20:0248
7
7
7
0:2584
20:1535
7
5
20:4216
20:0404
0:1380
20:0878
Resources
We also need the shifted Hankel matrix
2
1
1
2
3 6
6 0:6 0:2
Gð2Þ Gð3Þ Gð4Þ
6
1:5
6
7 6 1
Hs ð3; 3Þ 5 4 Gð3Þ Gð4Þ Gð5Þ 5 5 6
6 0:32 0:04
6
Gð4Þ Gð5Þ Gð6Þ
6
4 1
1:75
0:164 0:008
2
1:064
1
0:32 0:04
1
1:75
0:164 0:008
1
1:875
0:0828 0:0016
0:1216
0:1943
6 20:2599
0:3485
6
A 5 O1 Hs ðn; nÞ C 1 5 6
4 20:009736 0:06056
20:04398
20:7719
6 20:4768
6
B 5 first 2 columns of C 5 6
4 0:7110
20:1441
C 5 first 2 rows of O 5
D 5 Gð0Þ 5
0
0
0
0
21:0926
20:3560
20:2312
0:5896
0:1419
2
1:5
0:2094
1
1:75
3
0:164 0:008 7
7
7
1
1:875 7
7
0:0828 0:0016 7
7
7
1
1:9375 5
0:0416 0:0003
0:03386
3
0:1736 7
7
7
20:1247 5
0:1976
3
20:8644
20:9415 7
7
7
20:4946 5
0:0909
0:6558
0:6740 0:0689
20:9411 0:4206 0:1569
Resources
Antsaklis, P.J., Michel, A.N., 1997. Linear Systems. McGraw-Hill, New York.
Belanger, P.R., 1995. Control Engineering. A Modern Approach. Saunders, Fort Worth, TX.
Chen, C.T., 1984. Linear System Theory and Design. HRW, New York.
D’Azzo, J.J., Houpis, C.H., 1988. Linear Control System Analysis and Design. McGraw-Hill,
New York.
Delchamps, D.F., 1988. State Space and Input-Output Linear Systems. Springer-Verlag,
New York.
Gupta, S.C., Hasdorff, L., 1970. Fundamentals of Automatic Control. Wiley, New York.
Kailath, T., 1980. Linear Systems. Prentice Hall, Englewood Cliffs, NJ.
Moler, C.B., Van Loan, C.F., 1978. Nineteen dubious ways to calculate the exponential of
a matrix. SIAM Review 20, 801836.
Patel, R.V., Munro, N., 1982. Multivariable System Theory and Design. Pergamon Press.
Sinha, N.K., 1988. Control Systems. HRW, New York.
Skogestad, S., Postlethwaite, I., 2005. Multivariable Feedback Control: Analysis and
Design. Wiley, Chichester.
343
344
CHAPTER 8 Properties of StateSpace Models
PROBLEMS
8.1
Find the equilibrium state and the corresponding output for the system
x1 ðk 1 1Þ
5
x2 ðk 1 1Þ
0
20:5
1
20:1
x1 ðkÞ
1
x2 ðkÞ
yðkÞ 5 ½ 1
x1 ðkÞ
0
1
uðkÞ
1
x2 ðkÞ
when
(a) u(k) 5 0
(b) u(k) 5 1
8.2
A mechanical system has the statespace equations
x1 ðk 1 1Þ
5
x2 ðk 1 1Þ
0
20:5
1
2a1
x ðkÞ
0 1
x2 ðkÞ
yðkÞ 5 ½ 1
0
x1 ðkÞ
1
uðkÞ
1
x2 ðkÞ
where a1 is dependent on viscous friction.
(a) Using the results of Chapter 4, determine the range of the parameter
a1 for which the system is internally stable.
(b) Predict the dependence of the parameter a1 on viscous friction, and
use physical arguments to justify your prediction. (Hint: Friction dissipates energy and helps the system reach its equilibrium.)
8.3
Determine the internal stability and the input-output stability of the following linear systems:
x1 ðk 1 1Þ
(a)
x2 ðk 1 1Þ
5
0:1
0
1
0:2
yðkÞ 5 ½ 1
2
x1 ðk 1 1Þ
3
2
7 6
6
4 x2 ðk 1 1Þ 5 5 4
(b) x ðk 1 1Þ
3
y5½1
1
20:2
0
0
0
x1 ðkÞ
x2 ðkÞ
x1 ðkÞ
1
0
0:2
uðkÞ
x2 ðkÞ
0:2
0
32
x1 ðkÞ
3
2
1
76
7 6
0:1 54 x2 ðkÞ 5 1 4 0
21
1
x3 ðkÞ
3
x1 ðkÞ
7
6
0 4 x2 ðkÞ 5
x3 ðkÞ
1
0
2
0
3
7
0 5uðkÞ
1
Problems
0:1 0:3 x1 ðkÞ
0
x1 ðk 1 1Þ
5
1
uðkÞ
1 0:2 x2 ðkÞ
0:2
x2 ðk 1 1Þ
x ðkÞ
yðkÞ 5 ½ 1 1 1
x2 ðkÞ
(c)
2
3 2
x1 ðk 1 1Þ
20:1
(d) 4 x2 ðk 1 1Þ 5 5 4 0:1
0:3
x3 ðk 1 1Þ
y5½1
0
32
3 2
20:3
0
1
x1 ðkÞ
1
0:1 54 x2 ðkÞ 5 1 4 1
0
21
0
x3 ðkÞ
2
3
x1 ðkÞ
1 4 x2 ðkÞ 5
x3 ðkÞ
3
0
0 5uðkÞ
1
8.4
Determine the stable, marginally stable, and unstable modes for each of
the unstable systems presented in Problem 8.3.
8.5
Determine the controllability and stabilizability of the systems presented in
Problem 8.3.
8.6
Transform the following system to standard form for uncontrollable systems,
and use the transformed system to determine if it is stabilizable:
2
3 2
32
3 2
3
0:05 0:09 0:1
1 0
x1 ðkÞ
x1 ðk 1 1Þ
4 x2 ðk 1 1Þ 5 5 4 0:05
1:1
21 54 x2 ðkÞ 5 1 4 0 1 5uðkÞ
0:05 20:9
1
0 1
x3 ðk 1 1Þ
x3 ðkÞ
2
3
x1 ðkÞ
y 5 ½ 1 0 1 4 x2 ðkÞ 5
x3 ðkÞ
8.7
Transform the system to the standard form for unobservable systems, and
use the transformed system to determine if it is detectable:
20:2 20:08 x1 ðkÞ
1
x1 ðk 1 1Þ
5
1
uðkÞ
0:125
0
0
x2 ðk 1 1Þ
x2 ðkÞ
x ðkÞ
yðkÞ 5 ½ 1 0:8 1
x2 ðkÞ
8.8
Determine the controllability and stabilizability of the systems presented in
Problem 8.3 with the input matrices changed to the following:
(a) B 5 ½ 1 0 T
(b) B 5 ½ 1
(c) B 5 ½ 1
1 0 T
0 T
(d) B 5 ½ 1
0
1 T
345
346
CHAPTER 8 Properties of StateSpace Models
8.9
An engineer is designing a control system for a chemical process with
reagent concentration as the sole control variable. After determining that
the system is not controllable, why is it impossible for him to control all the
modes of the system by an innovative control scheme using the same control
variables? Explain, and suggest an alternative solution to the engineer’s
problem.
8.10
The engineer introduced in Problem 8.9 examines the chemical process
more carefully and discovers that all the uncontrollable modes with concentration as control variable are asymptotically stable with sufficiently fast
dynamics. Why is it possible for the engineer to design an acceptable controller with reagent concentration as the only control variable? If such a
design is possible, give reasons for the engineer to prefer it over a design
requiring additional control variables.
8.11
Determine the observability and detectability of the systems of
Problem 8.3.
8.12
Repeat Problem 8.11 with the following output matrices:
(a) C 5 ½ 0 1 (b) C 5 ½ 0 1 0 (c) C 5 ½ 1 0 (d) C 5 ½ 1 0
8.13
0
Consider the system
2
0
A54 0
20:005
1
0
20:11
3
0
1 5
20:7
2 3
0
b 5 4 0 5 c 5 0:5
1
1
0
d50
(a) Can we design a controller for the system that can influence all its modes?
(b) Can we design a controller for the system that can observe all its modes?
Justify your answers using the properties of the system.
8.14
Consider the system (A, B, C) and the family of systems (αA, βB, γC) with
each of (α, β, γ) nonzero.
(a) Show that, if λ is an eigenvalue of A with right eigenvector v and left
eigenvector wT, then αλ is an eigenvalue of αA with right eigenvector
v/α and left eigenvector wT/α.
(b) Show that (A, B) is controllable if and only if (αA, βB) is controllable
for any nonzero constants (α, β).
(c) Show that system (A, C) is observable if and only if (αA, γC) is
observable for any nonzero constants (α, γ).
8.15
Show that any system in controllable form is controllable.
Problems
8.16
Show that any system in observable form is observable.
8.17
The realization (A, B, C, D) with B 5 C 5 In, D 5 0n3n, allows us to use
the MATLAB commands zpk or tf to obtain the resolvent matrix of A.
(a) Show that the realization is minimal (no pole-zero cancellation) for
any state matrix A (i) using the rank test, and then (ii) using the eigenvector test.
(b) Obtain the resolvent matrix of Example 7.7 using the MATLAB commands zpk and tf.
8.18
Obtain statespace representations for the following linear systems:
(a) In controllable form
(b) In observable form
(c) In diagonal form
z 1 0:5
(i) GðzÞ 5 3
ðz 2 0:1Þðz 1 0:1Þ
(ii) GðzÞ 5 5
(iii) GðzÞ 5
(iv) GðzÞ 5
zðz 1 0:5Þ
ðz 2 0:1Þðz 1 0:1Þðz 1 0:8Þ
z2 ðz 1 0:5Þ
ðz 2 0:4Þðz 2 0:2Þðz 1 0:8Þ
z2
zðz 2 0:1Þ
2 0:9z 1 0:8
8.19
Obtain the controller form that corresponds to a renumbering of the state
variables of the controllable realization (also known as phase variable
form) from basic principles.
8.20
Obtain the transformation matrix to transform a system in phase variable
form to controller form. Prove that the transformation matrix will also
perform the reverse transformation.
8.21
Use the two steps of Section 8.5.4 to obtain a second observable realization from controller form. What is the transformation that will take this
form to the first observable realization of Section 8.5.4?
8.22
Show that the observable realization obtained from the phase variable
form realizes the same transfer function.
8.23
Show that the transfer functions of the following systems are identical, and
give a detailed explanation.
0
1
0
d50
A5
b5
cT 5 0 1
20:02 0:3
1
0:1 0
1
d50
A5
b5
cT 5 21 2
0 0:2
1
347
348
CHAPTER 8 Properties of StateSpace Models
8.24
Obtain a parallel realization for the transfer function matrix presented in
Example 8.15.
8.25
Find the poles and zeros of the following transfer function matrices:
2 z 2 0:1
1 3
6 ðz10:1Þ2
6
(a) GðzÞ 5 6
4
0
z 1 0:1 7
7
1 7
5
z 2 0:1
2
3
1
6
z 2 0:1 7
6
7
6
z 2 0:2
1 7
6
7
7
(b) GðzÞ 5 6
6 ðz 2 0:1Þðz 2 0:3Þ z 2 0:3 7
6
7
6
2 7
4
5
0
z 2 0:3
0
8.26
The linearized model of the horizontal plane motion of the INFANTE
autonomous underwater vehicle (AUV) of Problem 7.14 is given by
2 3 2
32 3 2
3
20:14 20:69 0:0
x1
0:056
x_1
4 x_2 5 5 4 20:19 20:048 0:0 54 x2 5 1 4 20:23 5u
0:0
1:0
0:0
0:0
x3
x_3
2 3
x1
1 0 0 4 5
y1
5
x2
0 1 0
y2
x3
where x1 is the sway speed, x2 is the yaw angle, x3 is the yaw rate, and u is
the rudder deflection. For the discrete-time system with a sampling period
T 5 50 ms as in Problem 7.14:
(a) Obtain the zeros of the system using Rosenbrock’s system matrix.
(b) Determine the system decoupling zeros by testing the system’s controllability and observability.
8.27
The terminal composition of a binary distillation column uses reflux and
steam flow as the control variables. The 2-input-2-output system is governed
by the transfer function
2
12:8e2s
6 16:7s 1 1
6
GðsÞ 5 6
6 6:6e27s
4
10:9s 1 1
3
218:9e23s
21:0s 1 1 7
7
7
23s 7
219:4e 5
14:4s 1 1
Computer Exercises
Find the discrete transfer function of the system with DAC, ADC, and a
sampling period of one time unit; then determine the poles and zeros of
the discrete-time system.
8.28
Show that the λi is an uncontrollable mode of the pair (A, B) if and only if
it is an unobservable mode of its dual (AT, BT).
COMPUTER EXERCISES
8.29
Write computer programs to simulate the second-order systems described
in Problem 8.3 for various initial conditions. Obtain state plane plots and
discuss your results, referring to the solutions presented in Examples 8.1
and 8.2.
8.30
Repeat Problem 8.23 using a computer-aided design (CAD) package.
Comment on any discrepancies between CAD results and solution by
hand.
8.31
Write a MATLAB function that determines the equilibrium state of the
system with the state matrix A, the input matrix B, and a constant input u
as input parameters.
8.32
Select a second-order state equation in diagonal form for which some
trajectories converge to the origin and others diverge. Simulate the system
using Simulink, and obtain plots for one diverging trajectory and one converging trajectory with suitable initial states.
8.33
Write a MATLAB program to find the Hankel realization of a given transfer
function and use it to obtain the results of Example 8.24.
349
CHAPTER
State Feedback Control
9
OBJECTIVES
After completing this chapter, the reader will be able to do the following:
1. Design state feedback control using pole placement.
2. Design servomechanisms using statespace models.
3. Analyze the behavior of multivariable zeros under state feedback.
4. Design state estimators (observers) for statespace models.
5. Design controllers using observer state feedback.
State variable feedback allows the flexible selection of linear system dynamics.
Often, not all state variables are available for feedback, and the remainder of the
state vector must be estimated. This chapter includes an analysis of state feedback
and its limitations. It also includes the design of state estimators for use when
some state variables are not available and the use of state estimates in feedback
control.
Throughout this chapter, we assume that the state vector x is n 3 1, the control
vector u is m 3 1, and the output vector y is l 3 1. We drop the subscript d for
discrete system matrices.
9.1 State and output feedback
State feedback involves the use of the state vector to compute the control action
for specified system dynamics. Figure 9.1 shows a linear system (A, B, C ) with
constant state feedback gain matrix K. Using the rules for matrix multiplication,
we deduce that the matrix K is m 3 n so that for a single-input system, K is a row
vector.
The equations for the linear system and the feedback control law are, respectively, given by the following equations:
xðk 1 1Þ 5 AxðkÞ 1 BuðkÞ
yðkÞ 5 CxðkÞ
uðkÞ 5 2KxðkÞ 1 vðkÞ
(9.1)
(9.2)
351
352
CHAPTER 9 State Feedback Control
v(k)
x(k+1)
u(k)
+
+
B
y(k)
x(k)
C
T
A
–K
FIGURE 9.1
Block diagram of constant state feedback control.
v(k)
u(k)
+
x(k+1)
+
B
y(k)
x(k)
T
C
A
–K
FIGURE 9.2
Block diagram of constant output feedback control.
The two equations can be combined to yield the closed-loop state equation
xðk 11Þ 5 AxðkÞ 1 B½2KxðkÞ 1 vðkÞ
5 ½A 2 BKxðkÞ 1 BvðkÞ
(9.3)
We define the closed-loop state matrix as
Acl 5 A 2 BK
(9.4)
and rewrite the closed-loop system statespace equations in the form
xðk 11Þ 5 Acl xðkÞ 1 BvðkÞ
yðkÞ 5 CxðkÞ
(9.5)
The dynamics of the closed-loop system depend on the eigenstructure (eigenvalues and eigenvectors) of the matrix Acl. Thus, the desired system dynamics
can be chosen with appropriate choice of the gain matrix K. Limitations of this
control scheme are addressed in the next section.
For many physical systems, it is either prohibitively costly or impossible to
measure all the state variables. The output measurements y must then be used
to obtain the control u as shown in Figure 9.2.
The feedback control for output feedback is
uðkÞ 5 2Ky yðkÞ 1 vðkÞ
5 2Ky CxðkÞ 1 vðkÞ
(9.6)
9.2 Pole placement
Substituting in the state equation gives the closed-loop system
xðk 1 1Þ 5 AxðkÞ 1 B½2Ky CxðkÞ 1 vðkÞ
5 ½A 2 BKy CxðkÞ 1 BvðkÞ
(9.7)
The corresponding state matrix is
Ay 5 A 2 BKy C
(9.8)
Intuitively, less can be accomplished using output feedback than state feedback because less information is used in constituting the control law. In addition,
the postmultiplication by the C matrix in (9.8) restricts the choice of closed-loop
dynamics. However, output feedback is a more general design problem because
state feedback is the special case where C is the identity matrix.
9.2 Pole placement
Using output or state feedback, the poles or eigenvalues of the system can be
assigned subject to system-dependent limitations. This is known as pole placement,
pole assignment, or pole allocation. We state the problem as follows.
DEFINITION 9.1: POLE PLACEMENT
Choose the gain matrix K or Ky to assign the system eigenvalues to an arbitrary set
{λi, i 5 1, . . . , n}.
The following theorem gives conditions that guarantee a solution to the poleplacement problem with state feedback.
THEOREM 9.1: STATE FEEDBACK
If the pair (A, B) is controllable, then there exists a feedback gain matrix K that arbitrarily
assigns the system poles to any set {λi, i 5 1, . . . , n}. Furthermore, if the pair (A, B) is
stabilizable, then the controllable modes can all be arbitrarily assigned.
PROOF
Necessity
If the system is not controllable, then by Theorem 8.4, the product wiTB is zero for some left
eigenvector wiT. Premultiplying the closed-loop state matrix by wiT gives
wTi Acl 5 wTi ðA 2 BKÞ 5 λi wTi
Hence, the ith eigenpair (eigenvalue and eigenvector) of A is unchanged by state feedback and cannot be arbitrarily assigned. Therefore, controllability is a necessary condition for
arbitrary pole assignment.
353
CHAPTER 9 State Feedback Control
Sufficiency
We first give the proof for the single-input (SI) case where b is a column matrix. For a controllable
pair (A, B), we can assume without loss of generality that the pair is in controllable form. We
rewrite (9.4) as
bkT 5 A 2 Acl
Substituting the controllable form matrices gives
0n21 3 1
2
#
k1
1
k2
?
kn
3
In21
0n21 3 1
4
5 5
2a0 2a1 . . . . . . 2an21
"
2
3
0n21 3 1
In21
24 5
2ad0 2ad1 . . . . . . 2adn21
354
That is,
2
3 2
3
0n21 3 n
0n21 3 n
4 5 5 4 5
ad0 2 a0 ad1 2 a1 . . . . . . adn21 2 an21
k1 k2 ? kn
Equating the last rows of the matrices yields
k1
k2
? kn 5 ad0 2 a0
ad1 2 a1
:::
::: adn21 2 an21 (9.9)
which is the control that yields the desired characteristic polynomial coefficients.
Sufficiency Proof 2
We now give a more general proof by contraposition—that is, we assume that the result is not
true and prove that the assumption is not true. So we assume that the eigenstructure of the
jth mode is unaffected by state feedback for any choice of the matrix K and prove that
the system is uncontrollable. In other words, we assume that the jth eigenpair is the same
for the open-loop and closed-loop systems. Then we have
n
o
wTj BK 5 wTj fA 2 Acl g 5 λj wTj 2 λj wTj 5 0T
Assuming K full rank, we have wjTB 5 0. By Theorem 8.4, the system is not controllable.
For a controllable SI system, the matrix (row vector) K has n entries with n
eigenvalues to be assigned, and the pole placement problem has a unique solution.
Clearly, with more inputs, the K matrix has more rows and consequently more
unknowns than the n equations dictated by the n specified eigenvalues. This freedom can be exploited to obtain a solution that has desirable properties in addition
to the specified eigenvalues. For example, the eigenvectors of the closed-loop
state matrix can be selected subject to constraints. There is a rich literature that
covers eigenstructure assignment, but it is beyond the scope of this text.
The sufficiency proof of Theorem 9.1 provides a method to obtain the feedback
matrix K to assign the poles of a controllable system for the SI case with the system
in controllable form. This approach is explored later. We first give a simple procedure
applicable to low-order systems.
9.2 Pole placement
PROCEDURE 9.1: POLE PLACEMENT BY EQUATING COEFFICIENTS
1. Evaluate the desired characteristic polynomial from the specified eigenvalues λdi,
i 5 1,. . . ,n using the expression
n
∆dc ðλÞ 5 L ðλ 2 λdi Þ
(9.10)
i51
2. Evaluate the closed-loop characteristic polynomial using the expression
detfλIn 2ðA 2 BKÞg
(9.11)
3. Equate the coefficients of the two polynomials to obtain n equations to be solved for
the entries of the matrix K.
EXAMPLE 9.1: POLE ASSIGNMENT
Assign the eigenvalues {0.3 6 j0.2} to the pair
0
3
A5
1
4
b5
0
1
Solution
For the given eigenvalues, the desired characteristic polynomial is
∆dc ðλÞ 5 ðλ 2 0:3 2 j0:2Þðλ 2 0:3 1 j0:2Þ 5 λ2 2 0:6λ 1 0:13
The closed-loop state matrix is
0
1
k1
2
1
4
0
1
5
3 2 k1 4 2 k2
A 2 bkT 5
0
3
k2
The closed-loop characteristic polynomial is
detfλIn 2 A 2 bkT g 5 det
λ
21
2 ð3 2 k1 Þ λ 2 ð4 2 k2 Þ
5 λ2 2 ð4 2 k2 Þλ 2 ð3 2 k1 Þ
Equating coefficients gives the two equations
1. 4 2 k2 5 0.6.k2 5 3.4
2. 23 1 k1 5 0.13.k1 5 3.13
that is,
kT 5 ½3:13; 3:4
Because the system is in controllable form, the same result can be obtained as the
coefficients of the open-loop characteristic polynomial minus those of the desired characteristic polynomial using (9.9).
355
CHAPTER 9 State Feedback Control
9.2.1 Pole placement by transformation to controllable form
1
0
60
6
6
6^
Tc21 5 Cc C 21 5 6
60
6
6
40
0
0
0
0
^
0
^
1
1
t2n
3
1
t2n 7
7
7
&
^ 7
7 b
? tn22;n 7
7
7
? tn21;n 5
1
t2n
t3n
?
a2
?
an21
a3
?
1
^
1
&
?
^
0
0
?
0
1
3
7
07
7
^7
7
7
05
0
(9.12)
Ab
?
?
2
system can be transformed into
Any controllable single-input-single-output (SISO)
controllable form using the transformation
2
a1
6
a
6 2
6
Tc 5 CCc21 5 b Ab . . . An21 b 6
6 ^
6
4 an21
356
...
An21 b
21
(9.13)
tnn
where C is the controllability matrix, the subscript c denotes the controllable form,
and the terms tjn, j 5 2, . . . , n, are given by
t2n 5 2an21
tj11;n 5 2
j21
X
an2i21 tj2i;n ;
j 5 2; . . . ; n 2 1
i50
The state feedback for a system in controllable form is
u 5 2kTc xc 5 2kTc ðTc21 xÞ
kT 5 ad0 2 a0
ad1 2 a1
...
. . . adn21 2 an21 Tc 21
(9.14)
(9.15)
We now have the following pole placement procedure.
PROCEDURE 9.2
1. Obtain the characteristic polynomial of the pair (A, B) using the Leverrier algorithm
described in Section 7.4.1.
2. Obtain the transformation matrix Tc21 using the coefficients of the polynomial from
step 1.
3. Obtain the desired characteristic polynomial coefficients from the given eigenvalues
using (9.10).
4. Compute the state feedback matrix using (9.15).
Procedure 9.2 requires the numerically troublesome inversion of the controllability matrix to obtain T. However, it does reveal an important characteristic
9.2 Pole placement
of state feedback. From (9.15), we observe that the feedback gains tend to
increase as the change from the open-loop to the closed-loop polynomial coefficients increases. Procedure 9.2, like Procedure 9.1, works well for low-order
systems but can be implemented more easily using a computer-aided design
(CAD) program.
EXAMPLE 9.2
Design a feedback controller for the pair
2
3
0:1 0 0:1
A 5 4 0 0:5 0:2 5
0:2 0 0:4
2
3
0:01
b54 0 5
0:005
to obtain the eigenvalues {0.1, 0.4 6 j 0.4}.
Solution
The characteristic polynomial of the state matrix is
λ3 2 λ2 1 0:27λ 2 0:01 i:e:;
The transformation matrix Tc21 is
2
3
2
0 0
1
10 1:5
6
7
6
1 5 3 103 4 0
1
Tc21 5 4 0 1
1
1
0:73
5
4
0:6
a2 5 21; a1 5 0:27; a0 5 20:01
321
7
1:3 5
2
0:1923
6
5 103 4 20:0577
1:9
0:0173
1:25
0:625
0:3125
20:3846
3
7
0:1154 5
0:1654
The desired characteristic polynomial is
λ3 2 0:9λ2 1 0:4λ 2 0:032 i:e:;
ad2 5 20:9; ad1 5 0:4; ad0 5 20:032
Hence, we have the feedback gain vector
kT 5 ad0 2 a0
ad1 2 a1
ad2 2 a2 Tc 21
5 20:032 1 0:01 0:4 2 0:27
2
0:1923
6
20:9 1 1 3 103 4 20:0577
0:0173
1:25
0:625
0:3125
20:3846
3
7
0:1154 5
0:1654
5 210 85 40
9.2.2 Pole placement using a matrix polynomial
The gain vector for pole placement can be expressed in terms of the desired
closed-loop characteristic polynomial. The expression, known as Ackermann’s
formula, is
kT 5 tT1 ∆dc ðAÞ
tT1
Tc21
(9.16)
∆dc ðλÞ
where
is the first row of the matrix
of (9.13) and
is the desired
closed-loop characteristic polynomial. From Theorem 9.1, we know that state
feedback can arbitrarily place the eigenvalues of the closed-loop system for any
357
358
CHAPTER 9 State Feedback Control
controllable pair (A, b). In addition, any controllable pair can be transformed into
controllable form (Ac, bc). By the Cayley-Hamilton theorem, the state matrix
satisfies its own characteristic polynomial ∆(λ) but not that corresponding to the
desired pole locations. That is,
∆c ðAÞ 5
n
X
ai Ai 5 0;
an 5 1
adi Ai 6¼ 0;
adn 5 1
i50
∆dc ðAÞ 5
n
X
i50
Subtracting and using the identity Ac 5 Tc21 ATc gives
Tc21 ∆dc ðAÞTc 5
n21
X
ðadi 2 ai ÞAic
(9.17)
i50
The matrix in controllable form possesses an interesting property, which
we use in this proof. If the matrix raised to power i, with i 5 1, 2, . . . , n 2 1, is
premultiplied by the first elementary vector
eT1 5 1
0Tn21 3 1
the result is the (i 1 1)th elementary vector—that is,
eT1 Aic 5 eTi11 ;
i 5 0; 1; . . . ; n 2 1
(9.18)
Premultiplying (9.17) by the elementary vector e1, and then using (9.18), we
obtain
eT1 Tc21 ∆dc ðAÞTc 5
n21
X
adi 2 ai eT1 Aic
i50
5
5
n21
X
i50
ad0
adi 2 ai eTi11
2 a0
ad1 2 a1
? adn21 2 an21
Using (9.15), we obtain
eT1 Tc21 ∆dc ðAÞTc 5 ad0 2 a0 ad1 2 a1
5 kTc 5 kT Tc
?
adn21 2 an21
Postmultiplying by Tc21 and observing that the first row of the inverse is tT1 5 eT1 Tc21 ,
we obtain Ackermann’s formula (9.16).
Minor modifications in Procedure 9.2 allow pole placement using Ackermann’s
formula. The formula requires the evaluation of the first row of the matrix Tc21
rather than the entire matrix. However, for low-order systems, it is often simpler to
evaluate the inverse and then use its first row. The following example demonstrates
pole placement using Ackermann’s formula.
9.2 Pole placement
EXAMPLE 9.3
Obtain the solution described in Example 9.2 using Ackermann’s formula.
Solution
The desired closed-loop characteristic polynomial is
∆dc ðλÞ 5 λ3 2 0:9λ2 1 0:4λ 2 0:032 i:e:;
ad2 5 20:9; ad1 5 0:4; ad0 5 20:032
The first row of the inverse transformation matrix is
2
0:1923
1:25
6
3
T
T 21
t1 5 e1 Tc 5 10 1 0 0 4 20:0577 0:625
0:0173
5 103 0:1923 1:25
0:3125
20:3846
3
7
0:1154 5
0:1654
20:3846
We use Ackermann’s formula to compute the gain vector
kT 5 tT1 ∆dc ðAÞ
5 tT1 fA3 2 0:9A2 1 0:4A 2 0:032I3 g
5 10
3
0:1923 1:25
2
23 4
20:3846 3 10
3
6 0 18
4 68 44 5
36 0 48
5 210 85 40
9.2.3 Choice of the closed-loop eigenvalues
Procedures 9.1 and 9.2 yield the feedback gain matrix once the closed-loop
eigenvalues have been arbitrarily selected. The desired locations of the eigenvalues are directly related to the desired transient response of the system. In this
context, considerations similar to those made in Section 6.6 can be applied.
However, the designer must take into account that poles associated with fast
modes will lead to high gains for the state feedback matrix and consequently to a
high control effort. High gains may also lead to performance degradation due
to nonlinear behavior such as ADC or actuator saturation. If all the desired closedloop eigenvalues are selected at the origin of the complex plane, the deadbeat
control strategy is implemented (see Section 6.7), and the closed-loop characteristic
polynomial is chosen as
∆dc ðλÞ 5 λn
(9.19)
Substituting in Ackermann’s formula (9.16) gives the feedback gain matrix
kT 5 tT1 An
(9.20)
The resulting control law will drive all the states to zero in at most n sampling
intervals starting from any initial condition. However, the limitations of deadbeat
control discussed in Section 6.7 apply—namely, the control variable can assume
unacceptably high values, and undesirable intersample oscillations can occur.
359
360
CHAPTER 9 State Feedback Control
EXAMPLE 9.4
Determine the gain vector k using Ackermann’s formula for the discretized statespace
model of the armature-controlled DC motor (see Example 7.16) where
2
3
0:1
0:0
0:9995
0:0095 5 and
20:0947 0:8954
1:0
Acl 5 4 0:0
0:0
2
3
1:622 3 1026
24
4
Bcl 5 4:821 3 10 5
9:468 3 1022
for the following choices of closed-loop eigenvalues:
1. {0.1, 0.4 6 j0.4}
2. {0.4, 0.6 6 j0.33}
3. {0, 0, 0} (deadbeat control)
Simulate the system in each case to obtain the zero-input response starting from the initial
condition x(0) 5 [1,1,1], and discuss the results.
Solution
The characteristic polynomial of the state matrix is
∆ðλÞ 5 λ3 2 2:895λ2 1 2:791λ 2 0:896
That is,
a2 5 22:895;
a1 5 2:791;
a0 5 20:896
The controllability matrix of the system is
2
23 4
C 5 10
3
0:001622 0:049832 0:187964
0:482100 1:381319 2:185571 5
94:6800 84:37082 75:73716
Using (9.13) gives the transformation matrix
2
Tc21
44
5 10
1:0527
21:9948
0:94309
20:0536
0:20688
20:14225
3
0:000255
20:00102 5
0:00176
1. The desired closed-loop characteristic polynomial is
∆dc ðλÞ 5 λ3 2 0:9λ2 1 0:4λ 2 0:032
That is,
ad2 5 20:9;
ad1 5 0:4;
ad0 5 20:032
By Ackermann’s formula, the gain vector is
kT 5 tT1 ∆dc ðAÞ
5 tT1 fA3 2 0:9A2 1 0:4A 2 0:032I3 g
5 104 ½ 1:0527
20:0536 0:000255 3 fA3 2 0:9A2 1 0:4A 2 0:032I3 g
5 103 ½ 4:9268 1:4324 0:0137 The zero-input response for the three states and the corresponding control variable u
are shown in Figure 9.3.
9.2 Pole placement
1.5
2
1
0
x2
x1
0.5
0
–0.5
–4
0
0.05
0.1
0.15
Time s
–6
0.2
500
0
10
1
0.05
0.1
0.15
Time s
0.2
0.05
0.1
0.15
Time s
0.2
4
0.5
u
x3
0
–500
–1000
–2
0
–0.5
0
0.05
0.1
0.15
Time s
0.2
–1
0
FIGURE 9.3
Zero-input state response and control variable for case 1 of Example 9.4.
2. The desired closed-loop characteristic polynomial is
∆dc ðλÞ 5 λ3 2 1:6λ2 1 0:9489λ 2 0:18756
That is,
ad2 5 21:6;
ad1 5 0:9489;
ad0 5 20:18756
And the gain vector is
kT 5 tT1 ∆dc ðAÞ
5 tT1 fA3 2 1:6A2 1 0:9489A 2 0:18756I3 g
5 104 1:0527 20:0536 0:000255 3 A3 2 1:6A2 1 0:9489A 2 0:18756I3 g
5 103 1:6985 0:70088 0:01008 The zero-input response for the three states and the corresponding control variable u
are shown in Figure 9.4.
3. The desired closed-loop characteristic polynomial is
∆dc ðλÞ 5 λ3
i:e:;
ad2 5 0;
ad1 5 0;
ad0 5 0
and the gain vector is
kT 5 tT1 ∆dc ðAÞ
5 tT1 fA3 g
5 104 ½ 1:0527 20:0536 0:000255 3 fA3 g
5 104 ½ 1:0527 0:2621 0:0017 The zero-input response for the three states and the corresponding control variable u
are shown in Figure 9.5.
361
1.5
1
1
0
0.5
–1
x2
x1
CHAPTER 9 State Feedback Control
0
–0.5
–2
0
0.05
0.1
0.15
Time s
2000
0
1000
0.05
0.1
0.15
Time s
0.05
0.1
0.15
Time s
0.2
0
u
–100
x3
–3
0
0.2
100
–1000
–200
–300
–2000
0
0.05
0.1
0.15
Time s
0.2
–3000
0
0.2
FIGURE 9.4
Zero-input state response and control variable for case 2 of Example 9.4.
2
1.5
0
x2
x1
1
–2
0.5
–4
0
0
0.05
0.1
0.15
Time s
–6
0.2
1500
3
1000
2
500
0.05
0.1
0.15
Time s
0.2
0.05
0.1
0.15
Time s
0.2
10 4
u
0
–500
–1
–1000
–1500
0
1
0
x3
362
0
0.05
0.1
0.15
Time s
0.2
–2
0
FIGURE 9.5
Zero-input state response and control variable for case 3 of Example 9.4.
9.2 Pole placement
We observe that when eigenvalues associated with faster modes are selected,
higher gains are required for the state feedback and the state variables have transient
responses with larger oscillations. Specifically, for the deadbeat control of case 3, the
gain values are one order of magnitude larger than those of cases 1 and 2, and the magnitude of its transient oscillations is much larger. Further, for deadbeat control, the
zero state is reached in n 5 3 sampling intervals, as predicted by the theory. However,
unpredictable transient intersample behavior occurs between 0.01 and 0.02 in the
motor velocity x2. This is shown in Figure 9.6, where the analog velocity and the sampled
velocity of the motor are plotted.
1
0
–1
–2
x2
–3
–4
–5
–6
–7
–8
–9
0
0.02 0.04 0.06 0.08
0.1 0.12 0.14 0.16
Time s
0.18
0.2
FIGURE 9.6
Sampled and analog velocity of the motor with deadbeat control.
Another consideration in selecting the desired closed-loop poles is the robustness of the system to modeling uncertainties. Although we have so far assumed
that the statespace model of the system is known perfectly, this is never true in
practice. Thus, it is desirable that control systems have poles with low sensitivity
to perturbations in their statespace matrices. It is well known that low pole sensitivity is associated with designs that minimize the condition number of the
modal matrix of eigenvectors of the state matrix.
For a SISO system, the eigenvectors are fixed once the eigenvalues are chosen.
Thus, the choice of eigenvalues determines the pole sensitivity of the system. For
example, selecting the same eigenvalues for the closed-loop system leads to a high
condition number of the eigenvector matrix and therefore to a closed-loop system
that is very sensitive to coefficient perturbations. For multi-input-multi-output (MIMO)
systems, the feedback gains for a given choice of eigenvalues are nonunique. This
363
364
CHAPTER 9 State Feedback Control
allows for more than one choice of eigenvectors and can be exploited to obtain robust
designs. However, robust pole placement for MIMO systems is beyond the scope of
this text and is not discussed further.
9.2.4 MATLAB commands for pole placement
The pole placement command is place. The following example illustrates the use
of the command:
.. A 5 ½0; 1; 3; 4;
.. B 5 ½0; 1;
.. poles 5 ½0:3 1 j :2; 0:3 2 j 0:2;
.. K 5 placeðA;B; polesÞ
place: ndigits 5 16
K 5 3:1300
3:4000
ndigits is a measure of the accuracy of pole placement.
It is also possible to compute the state feedback gain matrix of (9.15) or (9.16)
using basic MATLAB commands as follows:
1. Generate the characteristic polynomial of the state matrix to use (9.15).
.. polyðAÞ
2. Obtain the coefficients of the desired characteristic polynomial from a set of
desired eigenvalues given as the entries of vector poles.
.. desired 5 polyðpolesÞ
The vector desired contains the desired coefficients in descending order.
3. For (9.16), generate the polynomial matrix for a matrix A corresponding to
the desired polynomial:
.. polyvalmðdesired; AÞ
9.2.5 Pole placement for multi-input systems
For multi-input systems, the solution to the pole placement problem for a
specified set of eigenvalues is not unique. However, there are constraints on the
eigenvectors of the closed-loop matrix. Using the singular value decomposition
of the input matrix (see Appendix III), we can obtain a complete characterization of the eigenstructure of the closed-loop matrix that is assignable by state
feedback.
9.2 Pole placement
THEOREM 9.2
For any controllable pair (A, B), there exists a state feedback matrix K that assigns the eigenpairs
{(λi, vdi), i 5 1, . . . , n if and only if
U1T ½A 2 λi In vdi 5 0n 3 1
where
B5U
Z
(9.21)
0n2m 3 m
(9.22)
with
U 5 ½U0 jU1 ; U 21 5 U T
(9.23)
K 5 2Z 21 U0T ½Vd ΛVd21 2 A
(9.24)
The state feedback gain is given by
PROOF
The singular value decomposition of the input matrix is
Z
ΣB
VT 5 U
B5U
0n2m 3 m
0n2m 3 m
Z 5 ΣB V T
where the matrices of singular vectors satisfy U 21 5 UT, V 21 5 V T and the inverse of the
matrix Z is
Z 21 5 VΣ21
B
The closed-loop state matrix is
Acl 5 A 2 BK 5 Vd ΛVd21
with Λ 5 diagfλ1 ; . . . ; λn g; Vd 5 ½vd1 ; . . . ; vdn We write the matrix of left singular vectors U as in (9.23), and then multiply the closed-loop
state matrix
" T#
U0
½A 2 Vd ΛVd21 5 U T BK
U T ½A 2 Vd ΛVd21 5
U1T
Using the singular value decomposition of B, we have
" T
#
U0 A 2 Vd ΛVd21
ZK
5
0n2m 3 m
U1T A 2 Vd ΛVd21
The lower part of the matrix equality is equivalent to (9.21), while the upper part gives (9.24).
Condition (9.21) implies that the eigenvectors that can be assigned by state
feedback must be in the null space of the matrix U1T ; that is, they must be mapped to
zero by the matrix. This shows the importance of the singular value decomposition
365
366
CHAPTER 9 State Feedback Control
of the input matrix and its influence on the choice of available eigenvectors.
Recall that the input matrix must also be in a form that makes the pair (A, B)
controllable. While the discussion of this section helps us characterize state
feedback in the multivariable case, it does not really provide an easy recipe for
the choice of state feedback. In fact, no such recipe exists, although there are
alternative methods to choose the eigenvectors to endow the closed-loop system
with desirable properties, these methods are beyond the scope of this text.
EXAMPLE 9.5
For the pair
2
0
A54 0
2:005
1
0
2:11
3
0
1 5
20:7
2
0
B540
1
3
1
15
1
and the eigenvalues {0.1, 0.2, 0.3},
1. Determine a basis set for the space from which each eigenvector can be chosen.
2. Select a set of eigenvectors and determine the state feedback gain matrix that assigns
the desired eigenstructure. Verify that the closed-loop state matrix has the correct
eigenstructure.
Solution
The singular value decomposition of the input matrix is obtained using MATLAB
.. ½Ub; sigb; Vb 5 svdðBÞ
Ub 5
0:5000 0:5000 0:7071
0:5000 0:5000
0:7071
0:7071
0:7071
0
sigb 5
1:8478 0
0
0:7654
0
0
Vb 5
0:3827 0:9239
0:9239 0:38
The system is third order with two inputs—that
of left singular vectors is partitioned as
2
20:5
Ub 5 ½U0 jU1 4 20:5
20:7071
is, n 5 3, m 5 2, n 2 m 5 1— and the matrix
3
20:5 20:7071
20:5 0:7071 5
0:7071 0
For the eigenvalue λ1 5 0.1, equation (9.21) gives
U1T ½ A 2 0:1I3 vd1 5 ½ 0:0707
20:7778 0:7071 vd1 5 03 3 1
The corresponding eigenvector is
vd1 5 ½ 0:7383 0:4892 0:4643 T
9.3 Servo problem
Similarly, we determine the two remaining eigenvectors and obtain the modal matrix of
eigenvectors
2
3
0:7383 0:7620 0:7797
4
Vd 5 0:4892 0:4848 0:4848 5
0:4643 0:4293 0:3963
The closed-loop state matrix is
2
3
0:2377
2:0136
22:3405
21
4
1:0136
21:3405 5
Acl 5 Vd ΛVd 5 0:2377
0:6511 20:2695 20:6513
and the state feedback gain matrix is
K 5 2 Z 21 U0T ½Acl 2 A 5
20:4184
20:2377
22:3892
2:3405
1:1730
21:0136
The MATLAB command place gives the solution
Km 5
20:0346 0:3207
20:0499 0:0852
22:0996
0:7642
which also assigns the desired eigenvalues but with a different choice of eigenvectors.
9.2.6 Pole placement by output feedback
As one would expect, using output feedback limits our ability to assign the eigenvalues of the state system relative to what is achievable using state feedback. It
is, in general, not possible to arbitrarily assign the system poles even if the system
is completely controllable and completely observable. It is possible to arbitrarily
assign the controllable dynamics of the system using dynamic output feedback,
and a satisfactory solution can be obtained if the system is stabilizable and detectable. Several approaches are available for the design of such a dynamic controller.
One solution is to obtain an estimate of the state using the output and input of the
system and use it in state feedback as explained in Section 9.6.
9.3 Servo problem
The schemes shown in Figures 9.1 and 9.2 are regulators that drive the system state
to zero starting from any initial condition capable of rejecting impulse disturbances.
In practice, it is often necessary to track a constant reference input r with zero
steady-state error. For this purpose, a possible approach is to use the two degreeof-freedom control scheme in Figure 9.7, so called because we now have two
matrices to select: the feedback gain matrix K and the reference gain matrix F.
The reference input of (9.2) becomes v(k) 5 Fr(k), and the control law is
chosen as
uðkÞ 5 2KxðkÞ 1 FrðkÞ
(9.25)
367
368
CHAPTER 9 State Feedback Control
r(k)
v(k)
F
u(k)
+
x(k+1)
B
+
x(k)
T
y(k)
C
A
–K
FIGURE 9.7
Block diagram of the two degree-of-freedom controller.
with r(k) the reference input to be tracked. The corresponding closed-loop system
equations are
xðk 1 1Þ 5 Acl xðkÞ 1 BFrðkÞ
yðkÞ 5 CxðkÞ
(9.26)
where the closed-loop state matrix is
Acl 5 A 2 BK
The z-transform of the corresponding output is given by (see Section 7.8)
YðzÞ 5 C½zIn 2Acl 21 BFRðzÞ
The steady-state tracking error for a unit step input is given by
L im ðz 2 1ÞfYðzÞ 2 RðzÞg5 L imfC½zIn 2Acl 21 BF 2 Ig
z-1
z-1
5 C½In 2Acl 21 BF 2 I
For zero steady-state error, we require the condition
C½In 2Acl 21 BF 5 In
(9.27)
If the system is square (m 5 l) and Acl is stable (no unity eigenvalues), we solve
for the reference gain
F 5 ½CðIn 2Acl Þ21 B21
(9.28)
EXAMPLE 9.6
Design a statespace controller for the discretized statespace model of the DC motor
speed control system described in Example 6.9 (with T 5 0.02) to obtain zero steady-state
error due to a unit step, a damping ratio of 0.7, and a settling time of about 1 s.
Solution
The discretized transfer function of the system with digital-to-analog converter (DAC) and
analog-to-digital converter (ADC) is
GðsÞ
z 1 0:9293
5 1:8604 3 1024
GZAS ðzÞ 5 ð1 2 z21 Þ Z
s
ðz 2 0:8187Þðz 2 0:9802Þ
9.3 Servo problem
The corresponding statespace model, computed with MATLAB, is
x1 ðk 1 1Þ
1:799
5
x2 ðk 1 1Þ
1
x1 ðkÞ
0:01563
1
uðkÞ
x2 ðkÞ
0
x ðkÞ
yðkÞ 5 0:01191 0:01107 1
x2 ðkÞ
20:8025
0
The desired eigenvalues of the closed-loop system are selected as {0.9 6 j0.09}, as in
Example 6.9. This yields the feedback gain vector
K 5 20:068517 0:997197 and the closed-loop state matrix
Acl 5
20:8181
0
1:8
1
The feedforward gain is
F 5 C ðIn 2Acl Þ21 B
"
5
21
0:01191 0:01107
1:8
1 0
2
1
0 1
20:8181
0
21
0:01563
0
#21
5 50:42666
The response of the system to a step reference input r is shown in Figure 9.8. The system
has a settling time of about 0.84 s and percentage overshoot of about 4%, with a peak
time of about 1 s. All design specifications are met.
1.4
System: syscl
Peak amplitude: 1.04
Overshoot (%): 4.22
At time (sec): 0.64
1.2
System: syscl
Settling time (sec): 0.842
Amplitude
1
0.8
0.6
0.4
0.2
0
0
0.2
0.4
0.6
Time s
0.8
FIGURE 9.8
Step response of the closed-loop system of Example 9.6.
1
1.2
369
370
CHAPTER 9 State Feedback Control
r(k)
+
x(k + 1)
T
x(k)
v(k) u(k)
+
B
−K
x(k+1)
T
+
x(k)
y(k)
C
A
−K
−1
FIGURE 9.9
Control scheme with integral control.
The control law (9.25) is equivalent to a feedforward action determined by F to yield
zero steady-state error for a constant reference input r. Because the forward action does
not include any form of feedback, this approach is not robust to modeling uncertainties.
Thus, modeling errors (which always occur in practice) will result in nonzero steady-state
error. To eliminate such errors, we introduce the integral control shown in Figure 9.9, with
a new state added for each control error integrated.
The resulting statespace equations are
xðk 1 1Þ 5 AxðkÞ 1 BuðkÞ
xðk 1 1Þ 5 xðkÞ 1 rðkÞ 2 yðkÞ
(9.29)
yðkÞ 5 CxðkÞ
uðkÞ 5 KxðkÞ 2 KxðkÞ
where x is l 3 1. The statespace equations can be combined and rewritten in terms of an
augmented state vector xa ðkÞ 5 xðkÞ xðkÞ T as
xðk 1 1Þ
A
5
xðk 1 1Þ
2C
yðkÞ 5 C
0
Il
0
B
xðkÞ
2
K
xðkÞ
0
xðk 1 1Þ
xðk 1 1Þ
xðkÞ
0
1
rðkÞ
xðkÞ
Il
K
That is,
~ a ðkÞ 1
xa ðk 1 1Þ 5 ðA~ 2 B~KÞx
yðkÞ 5 C
0
Il
rðkÞ
(9.30)
0 xa ðkÞ
where
A~ 5
A
2C
0
Il
B
B~ 5
0
K~ 5 K
K
(9.31)
It can be shown that the system of (9.31) is controllable if and only if the original system
~ can
is controllable. The eigenvalues of the closed-loop system state matrix Acl 5 ðA~ 2 B~KÞ
be arbitrarily assigned by computing the gain matrix K~ using any of the procedures for the
regulator problem as described in Section 9.2.
9.3 Servo problem
EXAMPLE 9.7
Solve the design problem presented in Example 9.6 using integral control.
Solution
The statespace matrices of the system are
A5
20:8025
0
1:799
1
B5
0:01563
0
C 5 0:01191 0:01107 Adding integral control, we obtain
2
1:799
A
0
A~ 5
54
1
2C 1
20:01191
3
20:8025 0
0
05
20:01107 1
2
3
0:01563
B
5
B~ 5
54
0
0
0
In Example 9.6, the eigenvalues were selected as {0.9 6 j0.09}. Using integral control
increases the order of the system by one, and an additional eigenvalue must be selected. The
desired eigenvalues are selected as {0.9 6 j0.09, 0.2}, and the additional eigenvalue at 0.2 is
chosen for its negligible effect on the overall dynamics. This yields the feedback gain vector
K~ 5 51:1315
240:4431
The closed-loop system state matrix is
2
1
Acl 5 4 1
20:0119
240:3413
3
20:1706 0:6303
0
0 5
20:0111
1
The response of the system to a unit step reference signal r is shown in Figure 9.10.
The figure shows that the control specifications are satisfied. The settling time of 0.87 is
1.4
System: syscl
Peak amplitude: 1.04
Overshoot (%): 4.21
At time (sec): 0.66
1.2
System: syscl
Settling time (sec): 0.867
Amplitude
1
0.8
0.6
0.4
0.2
0
0
0.2
0.4
0.6
Time s
0.8
FIGURE 9.10
Step response of the closed-loop system of Example 9.7.
1
1.2
371
372
CHAPTER 9 State Feedback Control
well below the specified value of 1 s, and the percentage overshoot is about 4.2%, which
is less than the value corresponding to ζ 5 0.7 for the dominant pair.
9.4 Invariance of system zeros
A severe limitation of the state-feedback control scheme is that it cannot change
the location of the zeros of the system, which significantly affect the transient
response. To show this, we consider the z-transform of the system (9.5) including
the direct transmission matrix D:
2ðzI 2 A 1 BKÞXðzÞ 1 BVðzÞ 5 0
CXðzÞ 1 DVðzÞ 5 YðzÞ
(9.32)
If z 5 z0 is a zero of the system, then Y(z0) is zero with V(z0) and X(z0) nonzero.
Thus, for z 5 z0, the statespace equation (9.32) can be rewritten as
2ðz0 I 2 A 1 BKÞ
C
B
D
Xðz0 Þ
50
Vðz0 Þ
(9.33)
We rewrite (9.33) in terms of the statespace matrices of the open-loop
system and a vector V(z0) as
2ðz0 In 2 AÞ
C
B
D
2ðz0 In 2 A 1 BKÞ
Xðz0 Þ
5
Vðz0 Þ
C 2 DK
B
D
0
Xðz0 Þ
5
0
Vðz0 Þ 1 KXðz0 Þ
We observe that with the state feedback u(k) 5 2Kx(k) 1 r(k), the statespace
quadruple (A, B, C, D) becomes
ðA 2 BK; B; C 2 DK; DÞ
Thus, the zeros of the closed-loop system are the same as those of the plant and
are invariant under state feedback. Similar reasoning can establish the same result
for the system of (9.26).
EXAMPLE 9.8
Consider the following continuous-time system:
GðsÞ 5
s11
ð2s 1 1Þð3s 1 1Þ
Obtain a discrete model for the system with digital control and a sampling period T 5 0.02,
and then design a statespace controller with integral control and with the same closedloop eigenvalues as shown in Example 9.7.
9.4 Invariance of system zeros
Solution
The analog system with DAC and ADC has the transfer function
GðsÞ
z 2 0:9802
5 33:338 3 1024
GZAS ðzÞ 5 ð1 2 z21 Þ Z
s
ðz 2 0:9934Þðz 2 0:99Þ
with an open-loop zero at 0.9802. The corresponding statespace model (computed with
MATLAB) is
1:983 20:983 x1 ðkÞ
0:0625
x1 ðk 1 1Þ
5
1
uðkÞ
x2 ðk 1 1Þ
x2 ðkÞ
1
0
0
x ðkÞ
yðkÞ 5 ½ 0:0534 0:0524 1
x2 ðkÞ
The desired eigenvalues of the closed-loop system are selected as {0.9 6 j0.09} with an
additional eigenvalue at 0.2, and this yields the feedback gain vector
K~ 5 15:7345
The closed-loop system state matrix is
2
1
Acl 5 4
1
0:05342
224:5860
22:19016
3
20:55316 13:6885
5
0
0
20:05236
1
The response of the system to a unit step reference signal r, shown in Figure 9.11,
has a huge peak overshoot due to the closed-loop zero at 0.9802. The closed-loop control
cannot change the location of the zero.
4
3.5
3
Amplitude
2.5
2
1.5
1
0.5
0
0
0.5
1
Time s
FIGURE 9.11
Step response of the closed-loop system of Example 9.8.
1.5
373
374
CHAPTER 9 State Feedback Control
9.5 State estimation
In most applications, measuring the entire state vector is impossible or prohibitively expensive. To implement state feedback control, an estimate x^ ðkÞ of the
state vector can be used. The state vector can be estimated from the input and
output measurements by using a state estimator or observer.
9.5.1 Full-order observer
To estimate all the states of the system, one could in theory use a system with the
same state equation as the plant to be observed. In other words, one could use the
open-loop system
x^ ðk 11Þ 5 A^xðkÞ 1 BuðkÞ
However, this open-loop estimator assumes perfect knowledge of the system
dynamics and lacks the feedback needed to correct the errors that are inevitable in
any implementation. The limitations of this observer become obvious on examining
its error dynamics. We define the estimation error as x~ 5 x 2 x^ . We obtain the
error dynamics by subtracting the open-loop observer dynamics from the system
dynamics (9.1).
x~ ðk 11Þ 5 A~xðkÞ
The error dynamics are determined by the state matrix of the system and cannot
be chosen arbitrarily. For an unstable system, the observer will be unstable and
cannot track the state of the system.
A practical alternative is to feed back the difference between the measured
and the estimated output of the system, as shown in Figure 9.12. This yields to
the following observer:
x^ ðk 11Þ 5 A^xðkÞ 1 BuðkÞ 1 L½yðkÞ 2 C x^ ðkÞ
x̂ (k + 1)
B
u(k)
x(k + 1) = Ax(k) + Bu(k)
y(k) = Cx(k)
+
ˆ
x(k)
T
L
A
+
–C
y(k)
FIGURE 9.12
Block diagram of the full-order state estimator.
(9.34)
9.5 State estimation
Subtracting the observer state equation from the system dynamics yields the
estimation error dynamics
x~ ðk 1 1Þ 5 ðA 2 LCÞ~xðkÞ
(9.35)
The error dynamics are governed by the eigenvalues of the observer matrix
A0 5 A 2 LC. We transpose the matrix to obtain
AT0 5 AT 2 C T LT
(9.36)
which has the same eigenvalues as the observer matrix. We observe that (9.36) is
identical to the controller design equation (9.4) with the pair (A, B) replaced by
the pair (AT, CT). We therefore have Theorem 9.2.
THEOREM 9.3: STATE ESTIMATION
If the pair (A, C) is observable, then there exists a feedback gain matrix L that arbitrarily
assigns the observer poles to any set {λi, i 5 1, . . . , n}. Furthermore, if the pair (A, C) is
detectable, then the observable modes can all be arbitrarily assigned.
PROOF
Based on Theorem 8.12, the system (A, C) is observable (detectable) if and only if (AT, CT) is
controllable (stabilizable). Therefore, Theorem 9.3 follows from Theorem 9.1.
Based on Theorem 9.3, the matrix gain L can be determined from the desired
observer poles, as discussed in Section 9.2. Hence, we can arbitrarily select the
desired observer poles or the associated characteristic polynomial. From (9.32) it
follows that the MATLAB command for the solution of the observer pole placement problem is
.. L 5 placeðA0 ; C0 ; polesÞ0
EXAMPLE 9.9
Determine the observer gain matrix L for the discretized statespace model of the
armature-controlled DC motor of Example 9.4 with the observer eigenvalues selected as
{0.1, 0.2 6 j0.2}.
Solution
Recall that the system matrices are
2
3
2
3
1:0
0:1
0:0
1:622 3 1026
A 5 4 0:0
0:9995
0:0095 5 B 5 4 4:821 3 1024 5 C 5 1 0
0:0 20:0947 0:8954
9:468 3 1022
The MATLAB command place gives the observer gain
2
3
2:3949
4
L 5 18:6734 5
4:3621
0
375
376
CHAPTER 9 State Feedback Control
Expression (9.34) represents a prediction observer, because the estimated
state vector (and any associated control action) at a given sampling instant does
not depend on the current measured value of the system output. Alternatively, a
filtering observer estimates the state vector based on the current output (assuming negligible computation time), using the expression
x^ ðk 1 1Þ 5 A^xðkÞ 1 BuðkÞ 1 L½yðk 1 1Þ 2 CðA^xðkÞ 1 BuðkÞÞ
(9.37)
Clearly, the current output y(k 1 1) is compared to its estimate based on variables
at the previous sampling instant. The error dynamics are now represented by
x~ ðk 1 1Þ 5 ðA 2 LCAÞ~xðkÞ
(9.38)
Expression (9.38) is the same as Expression (9.35) with the matrix product CA
substituted for C. The observability matrix O of the system (A, CA) is
3
CA
6 7
7
6
6 CA2 7
7
6
75 OA
O56
7
6
6 ^ 7
7
6
4 5
CAn
2
where O is the observability matrix of the pair (A, C) (see Section 8.3). Thus,
if the pair (A, C) is observable, the pair (A, CA) is observable unless A has one
or more zero eigenvalues. If A has zero eigenvalues, the pair (A, CA) is
detectable because the zero eigenvalues are associated with stable modes. Further,
the zero eigenvalues are associated with the fastest modes, and the design of the
observer can be completed by selecting a matrix L that assigns suitable values
to the remaining eigenvalues of A 2 LCA.
EXAMPLE 9.10
Determine the filtering observer gain matrix L for the system of Example 9.9.
Solution
Using the MATLAB command place,
.. L 5 placeðA0 ; ðC AÞ0 ; polesÞ0
we obtain the observer gain
L5
1:0e 1 002
0:009910699530206
0:140383004697920
4:886483453094875
9.5 State estimation
9.5.2 Reduced-order observer
A full-order observer is designed so that the entire state vector is estimated from
knowledge of the input and output of the system. The reader might well ask, why
estimate n state variables when we already have l measurements that are linear
functions of the same variables? Would it be possible to estimate n 2 l variables
only and use them with measurement to estimate the entire state?
This is precisely what is done in the design of a reduced-order observer.
A reduced-order observer is generally more efficient than a full-order observer.
However, a full-order observer may be preferable in the presence of significant
measurement noise. In addition, the design of the reduced-order observer is more
complex.
We consider the linear time-invariant system (9.1) where the input matrix B
and the output matrix C are assumed full rank. Then the entries of the output
vector y(t) are linearly independent and form a partial state vector of length l,
leaving n 2 l variables to be determined. We thus have the state vector
2
3
3
C
y
4 5 5 Q21 x 5 4 5x
o
M
z
2
(9.39)
where M is a full-rank n 2 l 3 n matrix with rows that are linearly independent of
those of C, and z is the unknown partial state.
The statespace matrices for the transformed state variables are
h
Ct 5 CQo 5 Il
i
0l3n2l
2
3
B1 g l
4
5
Bt 5 Q21
o B 5 B2 g n 2 l
3
2
A1
A2 g l
7
6
At 5 Q21
o AQo 5 4 5
A3
A4 g n 2 l
|ffl{zffl} |ffl{zffl}
(9.40)
l
(9.41)
(9.42)
n2l
Thus, the state equation for the unknown partial state is
zðk 1 1Þ 5 A3 yðkÞ 1 A4 zðkÞ 1 B2 uðkÞ
(9.43)
We define an output variable to form a statespace model with (9.43) as
yz ðkÞ 5 yðk 1 1Þ 2 A1 yðkÞ 2 B1 uðkÞ 5 A2 zðkÞ
(9.44)
377
378
CHAPTER 9 State Feedback Control
This output represents the portion of the known partial state y(k 1 1) that is
computed using the unknown partial state. The observer dynamics, including the
error in computing yz, are assumed linear time invariant of the form
z^ ðk 1 1Þ 5 A3 yðkÞ 1 A4 z^ ðkÞ 1 B2 uðkÞ 1 L½yz ðkÞ 2 A2 z^ ðkÞ
5 ðA4 2 LA2 Þ^zðkÞ 1 A3 yðkÞ 1 L½yðk 1 1Þ 2 A1 yðkÞ 2 B1 uðkÞ 1 B2 uðkÞ
(9.45)
where z^ denotes the estimate of the partial state vector z. Unfortunately, the
observer (9.45) includes the term y(k 1 1), which is not available at time k. Moving
the term to the LHS reveals that its use can be avoided by estimating the variable
xðkÞ 5 z^ ðkÞ 2 LyðkÞ
(9.46)
Using (9.45) and the definition (9.46), we obtain the observer
xðk 1 1Þ 5 ðA4 2 LA2 ÞxðkÞ 1 ðA4 L 1 A3 2 LA1 2 LA2 LÞyðkÞ 1 ðB2 2 LB1 ÞuðkÞ
5 Ao xðkÞ 1 Ay yðkÞ 1 Bo uðkÞ
(9.47)
where
Ao 5 A4 2 LA2
Ay 5 Ao L 1 A3 2 LA1
(9.48)
Bo 5 B2 2 LB1
The block diagram of the reduced-order observer is shown in Figure 9.13.
The dynamic of the reduced-order observer (9.47) is governed by the matrix Ao.
The eigenvalues of Ao must be selected inside the unit circle and must be
sufficiently fast to track the state of the observed system. This reduces observer
design to the solution of (9.48) for the observer gain matrix L. Once L is obtained,
x ( k + 1)
x(k)
Bo
+
T
Ao
u(k)
x(k +1) = Ax(k) + Bu(k)
y(k) = Cx(k)
y(k)
Ay
L
ˆ
x(k)
+
To
FIGURE 9.13
Block diagram of the reduced-order observer.
9.5 State estimation
the other matrices in (9.47) can be computed and the state vector x^ can be obtained
using the equation
"
#
yðkÞ
x^ ðkÞ 5 Qo
z^ ðkÞ
(9.49)
#
"
#
"
#"
Il 0l3n2l
yðkÞ
yðkÞ
5 To
5 Qo
L
In2l
xðkÞ
xðkÞ
where the transformation matrix Qo is defined in (9.35).
Transposing (9.48) yields
T
T
ATo 5 A4 2 A2 LT
(9.50)
Because (9.50) is identical in form to the controller design equation (9.4), it can
be solved as discussed in Section 9.2. We recall that the poles of the matrix ATo
T
T
can be arbitrarily assigned provided that the pair ðA4 ; A2 Þ is controllable. From
the duality concept discussed in Section 8.6, this is equivalent to the observability
of the pair ðA4 ; A2 Þ. Theorem 9.4 gives a necessary and sufficient condition for
the observability of the pair.
THEOREM 9.4
The pair ðA4 ; A2 Þ is observable if and only if the system (A, C) is observable.
PROOF
The proof is left as an exercise.
EXAMPLE 9.11
Design a reduced-order observer for the discretized statespace model of the armaturecontrolled DC motor of Example 9.4 with the observer eigenvalues selected as {0.2 6 j0.2}.
Solution
Recall that the system matrices are
2
3
2
3
1:0
0:1
0:0
1:622 3 1026
24
A 5 4 0:0
0:9995
0:0095 5 B 5 4 4:821 3 10 5 C 5 1 0
0:0 20:0947 0:8954
9:468 3 1022
0
The output matrix C is in the required form, and there is no need for similarity transformation. The second and third state variables must be estimated. The state matrix is
partitioned as
3
2
2
3
1:0
0:1
0:0
aT2
a1
6 7
7
6
7 6
A 5 4 5 5 6 0:0
0:9995 0:0095 7
5
4
a3
A4
0:0
20:0947 0:8954
379
380
CHAPTER 9 State Feedback Control
The similarity transformation can be selected as an identity matrix; that is,
3
2
1
0
0
6 7
7
6
Q21
o 54 0
1
0 5
0
0
1
and therefore we have At 5 A, Bt 5 B, and Ct 5 C. Hence, we need to solve the linear
equation
0:9995
0:0095
2 l 0:1 0
Ao 5
20:0947 0:8954
to obtain the observer gain
l 5 14:949 550:191
T
The corresponding observer matrices are
20:4954 0:0095
255:1138 0:8954
14:949
0:04579
4:821 3 1024
26
2
3
1:622
3
10
bo 5 b2 2 lb1 5
5
3 1022
550:191
9:37876
9:468 3 1022
Ao 5
ay 5 Ao l 1 a3 2 la1
20:4954 0:0095
14:949
0
14:949
20:1713
5
1
2
315
3 102
255:1138 0:8954 550:191
0
550:191
28:8145
The state estimate can be computed using
x^ ðkÞ 5 Qo
1
l
0T1 3 2
I2
3
1
0T1 3 2
yðkÞ
yðkÞ
6
7
54 5
14:949
xðkÞ
xðkÞ
I2
550:191
2
9.6 Observer state feedback
If the state vector is not available for feedback control, a state estimator can be
used to generate the control action as shown in Figure 9.14. The corresponding
control vector is
uðkÞ 5 2K x^ ðkÞ 1 vðkÞ
(9.51)
Substituting in the state equation (9.1) gives
xðk 1 1Þ 5 AxðkÞ 2 BK x^ ðkÞ 1 BvðkÞ
(9.52)
Adding and subtracting the term BKx(k), we rewrite (9.52) in terms of the
estimation error x~ 5 x 2 x^ as
xðk 1 1Þ 5 ðA 2 BKÞxðkÞ 1 BK x~ ðkÞ 1 BvðkÞ
(9.53)
9.6 Observer state feedback
r(k)
u(k)
x(k)
x(k+1)
+
B
y(k)
T
+
C
A
xˆ ( k+1)
B
x̂(k)
−C
T
+
+
A
L
−K
FIGURE 9.14
Block diagram of a system with observer state feedback.
If a full-order (predictor) observer is used, by combining (9.53) with (9.35),
we obtain
xðk 1 1Þ
A 2 BK
5
x~ ðk 1 1Þ
0
BK
A 2 LC
xðkÞ
B
1
vðkÞ
x~ ðkÞ
0
(9.54)
The state matrix of (9.54) is block triangular and its characteristic polynomial is
∆cl ðλÞ 5 det½λI 2ðA 2 BKÞdet½λI 2ðA 2 LCÞ
5 ∆c ðλÞ∆o ðλÞ
(9.55)
Thus, the eigenvalues of the closed-loop system can be selected separately from
those of the observer. This important result is known as the separation theorem
or the uncertainty equivalence principle.
Analogously, if a reduced-order observer is employed, the estimation error x~
can be expressed in terms of the errors in estimating y and z as
2
3
y~ ðkÞ
x~ ðkÞ 5 Qo 4 5
z~ ðkÞ
y~ ðkÞ 5 yðkÞ 2 y^ ðkÞ z~ ðkÞ 5 zðkÞ 2 z^ ðkÞ
(9.56)
381
CHAPTER 9 State Feedback Control
We partition the matrix Qo into an n 3 l matrix Qy and an n 3 n 2 l matrix Qz to
allow the separation of the two error terms and rewrite the estimation error as
2
x~ ðkÞ 5 Qy
3
y~ ðkÞ
Qz 4 55 Qy y~ ðkÞ 1 Qz z~ ðkÞ
z~ ðkÞ
(9.57)
Assuming negligible measurement error y~ , the estimation error reduces to
x~ ðkÞ 5 Qz z~ ðkÞ
(9.58)
Substituting from (9.58) into the closed-loop equation (9.53) gives
xðk 1 1Þ 5 ðA 2 BKÞxðkÞ 1 BKQz z~ ðkÞ 1 BvðkÞ
(9.59)
Evaluating z~ ðkÞ by subtracting (9.45) from (9.43) and substituting A2 zðkÞ for yz, as
in (9.44), we obtain
z~ ðk 1 1Þ 5 ðA4 2 LA2 Þ~zðkÞ
(9.60)
Combining (9.59) and (9.60), we obtain the equation
xðk 1 1Þ
A 2 BK
5
0n2l 3 n
z~ ðk 1 1Þ
BKQz
A4 2 LA2
xðkÞ
B
1
vðkÞ
z~ ðkÞ
0
(9.61)
The state matrix of (9.61) is block triangular and its characteristic polynomial is
det½λI 2ðA4 2 LA2 Þdet½λI 2ðA 2 BKÞ
(9.62)
Thus, as for the full-order observer, the closed-loop eigenvalues for the reduced-order
observer state feedback can be selected separately from those of the reduced-order
observer. The separation theorem therefore applies for reduced-order observers as
well as for full-order observers.
In addition, combining the plant state equation and the estimator (9.47) and
using the output equation, we have
A
xðk 1 1Þ
5
Ay C
xðk 1 1Þ
0n 3 n2l
Ao
B
xðkÞ
1
uðkÞ
xðkÞ
B0
(9.63)
We express the estimator state feedback of (9.51) as
uðkÞ 5 2K x^ ðkÞ 1 vðkÞ
yðkÞ
1 vðkÞ 5 2K Toy
5 2KTo
xðkÞ
382
Tox
CxðkÞ
1 vðkÞ
xðkÞ
(9.64)
9.6 Observer state feedback
where Toy and Tox are partitions of To of (9.49) of order n 3 l and n 3 n 2 l,
respectively. Substituting in (9.63), we have
A 2 BKToy C
2BKTox
B
xðkÞ
xðk 1 1Þ
5
1
vðkÞ
(9.65)
Ay C 2 Bo KToy C Ao 2 Bo KTox xðkÞ
xðk 1 1Þ
B0
Equation (9.65) can be used to simulate the complete estimator state feedback system.
9.6.1 Choice of observer eigenvalues
In the selection of the observer poles or the associated characteristic polynomial,
expressions (9.55) or (9.62) must be considered. The choice of observer poles
is not based on the constraints related to the control effort discussed in
Section 9.2.3. However, the response of the closed-loop system must be dominated by the poles of the controller that meet the performance specifications.
Therefore, as a rule of thumb, the poles of the observer should be selected from
3 to 10 times faster than the poles of the controller. An upper bound on the speed
of response of the observer is imposed by the presence of the unavoidable
measurement noise. Inappropriately fast observer dynamics will result in tracking
the noise rather than the actual state of the system. Hence, a deadbeat observer,
although appealing in theory, is avoided in practice.
The choice of observer poles is also governed by the same considerations
related to the robustness of the system discussed in Section 9.2.3 for the state
feedback control. Thus, the sensitivity of the eigenvalues to perturbations in the
system matrices must be considered in the selection of the observer poles.
We emphasize that the selection of the observer poles does not influence the
performance of the overall control system if the initial conditions are estimated
perfectly. We prove this fact for the full-order observer. However, the result also
holds for the reduced-order observer, but the proof is left as an exercise for the
reader. To demonstrate this fact, we consider the state equation (9.54) with the
output equation (9.1):
xðk 1 1Þ
A 2 BK
5
0
x~ ðk 1 1Þ
yðkÞ 5 C
0
BK
A 2 LC
xðkÞ
x~ ðkÞ
xðkÞ
B
1
vðkÞ
x~ ðkÞ
0
(9.66)
The zero input-output response of the system (v(k) 5 0) can be determined iteratively as
yð0Þ 5 Cxð0Þ
yð1Þ 5 ½ C
0
A 2 BK
0
BK
A 2 LC
xð0Þ
5 CðA 2 BKÞxð0Þ 1 CBK x~ ð0Þ
x~ ð0Þ
383
384
CHAPTER 9 State Feedback Control
yð2Þ 5 C
0
A 2 BK
0
BK
A 2 LC
x ð1Þ
5 CðA 2 BK Þx ð1Þ 1 CBK x~ ð1Þ
x~ ð1Þ
5 C ðA2BK Þ2 x ð0Þ 2 C ðA 2 BK ÞBK x~ ð0Þ 2 CBK ðA 2 LC Þ~x ð0Þ
^
Clearly, the observer matrix L influences the transient response if and only if
x~ ð0Þ 6¼ 0. This fact is confirmed by the determination of the z-transfer function
from (9.66), which implicitly assumes zero initial conditions
21
B
A2BK
BK
5 CðzI2A1BKÞ21 B
GðzÞ 5 C 0 zI2
0
0
A2LC
where the observer gain matrix L does not appear.
Another important transfer function relates the output y(k) to the control u(k).
It is obtained from the observer state equation (9.34) with (9.51) as the output equation
and is given by
Gco ðzÞ 5 2 KðzIn 2A1BK 1LCÞ21 L
Note that any realization of this z-transfer function can be used to implement the
controller-observer regardless of the realization of the system used to obtain the
transfer function. The denominator of the transfer function is
detðzIn 2 A 1 BK 1 LCÞ
which is clearly different from (9.62). This should come as no surprise, since the
transfer function represents the feedback path in Figure 9.14 and not the closedloop system system dynamics.
EXAMPLE 9.12
Consider the armature-controlled DC motor of Example 9.4. Let the true initial condition
be x(0) 5 [1, 1, 1], and let its estimate be the zero vector x^ ð0Þ 5 ½0; 0; 0T . Design a full-order
observer state feedback for a zero-input response with a settling time of 0.2 s.
Solution
As Example 9.4 showed, a choice for the control system eigenvalues that meets the design
specification is {0.6, 0.4 6 j0.33}. This yields the gain vector
K 5 103 1:6985
20:70088 0:01008
The observer eigenvalues must be selected so that the associated modes are sufficiently
faster than those of the controller. We select the eigenvalues {0.1, 0.1 6 j0.1}. This yields the
observer gain vector
2
3
0:02595
24
L 5 10 0:21663 5
5:35718
9.6 Observer state feedback
1
5
1000
0
0
x
3
x
–5
x
1
2
0
–1000
–1
–10
–2
–15
–2000
0
0.1
0.2
Time s
0
0.1
0.2
Time s
10
1
0
0.1
0.2
Time s
0
0.1
0.2
Time s
500
0
6
0
x
x
x
4
5
0
–10
–1
–20
0
0.1
0.2
Time s
–500
0
0.1
0.2
Time s
FIGURE 9.15
Zero-input response for Example 9.12.
Using (9.66), we obtain the spacespace equations
2 0:9972
6
6 20:8188
6
6
6 2160:813
xðk 1 1Þ
56
6
x~ ðk 1 1Þ
6
6
6
4
0:0989
0
0:6616
0:0046
266:454
20:0589
0:0028
0:0011
0:3379 0:0049
2535:79
0 0
3
7
7
7
7
160:813 66:359 0:9543 7 xðkÞ
7
21:5949
0:1
0 7
7 x~ ðkÞ
7
7
221:663
0:9995
0:0095 5
0:8188
0
h
yðkÞ 5 1
0
i
0
xðkÞ
x~ ðkÞ
20:0947 0:8954
The response to the initial condition [1, 1, 1, 1, 1, 1] is plotted in Figure 9.15. We compare the plant state variables xi, i 5 1, 2, 3 to the estimation errors xi, i 5 4, 5, 6. We
observe that the estimation errors decay to zero faster than the system states and that the
system has an overall settling time less than 0.2 s.
385
386
CHAPTER 9 State Feedback Control
EXAMPLE 9.13
Solve Example 9.12 using a reduced-order observer.
Solution
In this case, we have l 5 1, and because the measured output corresponds to the first
element of the state we do not need similarity transformation, that is, Q0 5 Q021 5 I3. Thus,
we obtain
A1 5 1
A2 5 0:1
0
0
A3 5
B1 5 1:622 3 1026
0:9995
0:0095
20:0947 0:8954
A4 5
B2 5
4:821 3 1024
9:468 3 1022
We select the reduced-order observer eigenvalues as {0.1 6 j0.1} and obtain the observer
gain vector
L 5 102
0:16949
6:75538
and the associated matrices
Ao 5
20:6954 0:0095
267:6485 0:8954
Ay 5 103
20:02232
21:21724
Bo 5 1022
0:045461
9:358428
Partitioning Q0 5 To 5 I3 gives
2
0
Qz 5 4 1
0
2
3
3
0:01
0 0
24
4
5
Tox 5 1 0
Toy 5 10 0:16949 5
6:75538
0 1
3
0
05
1
2
We have that the statespace equation of (9.61) is
2
0:99725
6
6 20:81884
6
6
xðk 1 1Þ
56
6 2160:813
z~ðk 1 1Þ
6
6
4
0:09886
0:00002
0:00114
0:66161
0:00464
0:33789
266:4540
20:05885
0
0:00002
3
20:6954
7
0:00486 7
7
7 xðkÞ
0:95435 7
7 z~ ðkÞ
7
0:0095 7
5
267:6485
0:8954
66:3593
whereas the statespace equation of (9.65) is
2 0:96693
0:1
0
6
29:82821
0:9995
0:0095
6
6
xðk 1 1Þ
6
5 6 21930:17 20:0947 0:8954
xðk 1 1Þ
6
6
231:5855 0 0
4
23125:07 0 0
20:00002 3
7
20:33789 20:00486 7
7
7 xðkÞ
266:3593 20:95425 7
7 xðkÞ
7
21:01403 0:00492 5
20:00114
2133:240 0:04781
(9.67)
9.6 Observer state feedback
1.5
2
1
0
200
100
x3
x2
x1
0
0.5
–2
–100
0
–200
–4
–0.5
0
0.1
Time s
0.2
0
0.1
Time s
0.2
1
–300
0
0.1
Time s
0.2
0
0.5
x5
x4
–20
0
–40
–0.5
–60
0
0.1
Time s
0.2
0
0.1
Time s
0.2
FIGURE 9.16
Zero-input response for Example 9.13.
The response of the statespace system (9.67) to the initial condition [1, 1, 1, 1, 1] is
plotted in Figure 9.16. As in Example 9.12, we observe that the estimation errors xi, i 5 4,
and 5 decay to zero faster than the system states xi, i 5 1, 2, and 3, and that the system
has an overall settling time less than 0.2 s.
EXAMPLE 9.14
Consider the armature-controlled DC motor of Example 9.4. Design a full-order observer
state feedback with feedforward action for a step response with a settling time of 0.2 s
with an overshoot less than 10%.
Solution
As in Example 9.12, we select the control system eigenvalues as {0.4, 0.6 6 j0.33} and the
observer eigenvalues as {0.1, 0.1 6 j0.1}. This yields the gain vectors
2
K 5 10 1:6985
3
20:70088 0:01008 ;
L 5 10
24
3
0:02595
0:21663 5
5:35718
387
CHAPTER 9 State Feedback Control
System: syscl
Peak amplitude: 1.06
Overshoot (%): 6.45
At time (sec): 0.08
1.2
System: syscl
Settling time (sec): 0.112
1
Amplitude
388
0.8
0.6
0.4
0.2
0
0
0.02
0.04
0.06
0.08 0.1
Time s
0.12
0.14
0.16
0.18
0.2
FIGURE 9.17
Step response for Example 9.14.
From the matrix Acl 5 A 2 BK, we determine the feedforward term using (9.28) as
F 5 ½CðIn 2ðA2BKÞÞ21 B21 5 1698:49
Substituting Fr(k) for v(k) (9.66), we obtain the closed-loop statespace equations
0.9972
0.0989
0
0.0028
0.0011
0
–0.8188
0.6616
0.0046
0.8188
0.3379
0.0049
–66.454
–0.0589
160.813
66.359
0.9543
–160.813
x(k + 1)
=
~
x(k + 1)
–1.5949
0
0.1
0
–21.663
0.9995
0.0095
–535.79
–0.0947
0.8954
0.0028
0.8188
160.813
x(k)
+
~
x(k)
r(k)
0
y(k) = [1 0 0 | 0 ]
x(k)
~
x(k)
The step response of the system shown in Figure 9.17 has a settling time of about 0.1
and a percentage overshoot of about 6%. The controller meets the design specification.
9.7 Pole assignment using transfer functions
9.7 Pole assignment using transfer functions
The pole assignment problem can be solved in the framework of transfer functions. Consider the statespace equations of the two-degree-of-freedom controller
shown in Figure 9.7 with the state vector estimated using a full-order observer.
For a SISO plant with observer state feedback, we have
x^ ðk 1 1Þ 5 A^xðkÞ 1 BuðkÞ 1 LðyðkÞ 2 C x^ ðkÞÞ
uðkÞ 5 2K x^ ðkÞ 1 FrðkÞ
or equivalently,
x^ ðk 1 1Þ 5 ðA 2 BK 2 LCÞ^xðkÞ 1 ½ BF
uðkÞ 5 2K x^ ðkÞ 1 FrðkÞ
L
rðkÞ
yðkÞ
The corresponding z-transfer function from [r, y] to u is
UðzÞ 5 ½2KðzI2A1BK 1LCÞ21 BF 1 F RðzÞ 2 ½KðzI2A1BK 1LCÞ21 LYðzÞ
Thus, the full-order observer state feedback is equivalent to the transfer function
model depicted in Figure 9.18. In the figure, the plant G(z) 5 P(z)/Q(z) is assumed
strictly realizable; that is, the degree of P(z) is less than the degree of Q(z). We
also assume that P(z) and Q(z) are coprime (i.e., they have no common factors).
Further, it is assumed that Q(z) is monic (i.e., the coefficient of the term with the
highest power in z is one).
From the block diagram shown in Figure 9.18, simple block diagram manipulations give the closed-loop transfer function
YðzÞ
PðzÞNðzÞ
5
RðzÞ QðzÞDðzÞ 1 PðzÞSðzÞ
and the polynomial equation
ðQðzÞDðzÞ 1 PðzÞSðzÞÞYðzÞ 5 PðzÞNðzÞRðzÞ
(9.68)
Therefore, the closed-loop characteristic polynomial is
∆cl ðzÞ 5 QðzÞDðzÞ 1 PðzÞSðzÞ
R(z)
N(z)
D(z)
U(z)
+
P(z)
Q(z)
S(z)
D(z)
FIGURE 9.18
Block diagram for pole assignment with transfer functions.
(9.69)
Y(z)
389
390
CHAPTER 9 State Feedback Control
The pole placement problem thus reduces to finding polynomials D(z) and S(z)
that satisfy (9.69) for given P(z), Q(z), and for a given desired characteristic polynomial ∆cl(z). Equation (9.69) is called a Diophantine equation, and its solution
can be found by first expanding its RHS terms as
PðzÞ 5 pn21 zn21 1 pn22 zn22 1 . . . 1 p1 z 1 p0
QðzÞ 5 zn 1 qn21 zn21 1 . . . 1 q1 z 1 q0
DðzÞ 5 dm zm 1 dm21 zm21 1 . . . 1 d1 z 1 d0
SðzÞ 5 sm zm 1 sm21 zm21 1 . . . 1 s1 z 1 s0
The closed-loop characteristic polynomial ∆cl(z) is of degree n 1 m and has
the form
∆cl ðzÞ 5 zn1m 1 δn1m21 zn1m21 1 . . . 1 δ1 z 1 δ0
Thus, (9.69) can be rewritten as
zn1m 1 δn1m21 zn1m21 1 . . . 1 δ1 z 1 δ0 5 ðzn 1 qn21 zn21 1 . . . 1 q1 z 1 q0 Þ
ðdm zm 1 dm21 zm21 1 . . . 1 d1 z 1 d0 Þ 1 ðpn21 zn21 1 pn22 zn22 1 . . . 1 p1 z 1 p0 Þ
ðsm zm 1 sm21 zm21 1 . . . 1 s1 z 1 s0 Þ
(9.70)
Equation (9.70) is linear in the 2m unknowns di and si, i 5 0, 1, 2, . . . , m 2 1,
and its LHS is a known polynomial with n 1 m 2 1 coefficients. The solution of
the Diophantine equation is unique if n 1 m 2 1 5 2m—that is, if m 5 n 2 1.
Equation (9.70) can be written in the matrix form
2
1
0
1
0
0
6q
6 n21
6
6 qn22 qn21 1
6
6
^
^
6 ^
6
6 q0
q1 q2
6
6
q 0 q1
6 0
6
6 0
0 q0
6
6
^
^
4 ^
0
0
0
?
?
0
0
pn21
?
0
^
pn22 pn21 ?
^
^
0
0
0
?
?
? qn21
p0
p1
?
? qn22
? qn23
0
0
p0
0
?
?
^
0
^
0
?
?
^
q0
32
3 2
3
1
dm
0
7 6
7
6
0 7
76 dm21 7 6 δ2n22 7
76
7 6
7
7 6
7
6
0 7
76 dm22 7 6 δ2n23 7
76
7 6
7
^ 76 ^ 7 6 ^ 7
76
7 6
7
7 6
7
6
pn21 7
76 d0 7 5 6 δn21 7 (9.71)
76
7 6
7
pn22 76 sm 7 6 δn22 7
76
7 6
7
7 6
7
6
pn23 7
76 sm21 7 6 δn23 7
76
7 6
7
^ 54 ^ 5 4 ^ 5
p0
s0
δ0
It can be shown that the matrix on the LHS is nonsingular if and only if the
polynomials P(z) and Q(z) are coprime, which we assume. As discussed in
Section 9.2.3, the matrix must have a small condition number for the system to
be robust with respect to errors in the known parameters. The condition number
becomes larger as the matrix becomes almost singular.
The structure of the matrix shows that it will be almost singular if the coefficients of the numerator polynomial P(z) and denominator polynomial Q(z) are
9.7 Pole assignment using transfer functions
almost identical. We therefore require that the roots of the polynomials P(z) and
Q(z) be sufficiently different to avoid an ill-conditioned matrix. From the discusion of pole-zero matching of Section 6.3.2, it can be deduced that the poles
of the discretized plant approach the zeros as the sampling interval is reduced
(see also Section 12.2.2). Thus, when the controller is designed by pole assignment, to avoid an ill-conditioned matrix in (9.71), the sampling interval must not
be excessively short.
We now discuss the choice of the desired characteristic polynomial. From the
equivalence of the transfer function design to the statespace design described in
Section 9.6, the separation principle implies that ∆dcl ðzÞ can be written as the
product
∆dcl ðzÞ 5 ∆dc ðzÞ∆do ðzÞ
where ∆dc ðzÞ is the controller characteristic polynomial and ∆do ðzÞ is the observer
characteristic polynomial. We select the polynomial N(z) as
NðzÞ 5 kff ∆do ðzÞ
(9.72)
∆do ðzÞ
so that the observer polynomial
cancels in the transfer function from the
reference input to the system output. The scalar constant kff is selected so that the
steady-state output is equal to the constant reference input
Yð1Þ
Pð1ÞNð1Þ
Pð1Þkff ∆do ð1Þ
5 d
5
51
Rð1Þ
∆c ð1Þ∆do ð1Þ
∆dc ð1Þ∆do ð1Þ
The condition for zero steady-state error is
kff 5
∆dc ð1Þ
Pð1Þ
(9.73)
EXAMPLE 9.15
Solve Example 9.14 using the transfer function approach.
Solution
The plant transfer function is
GðzÞ 5
PðzÞ
1:622z2 1 45:14z 1 48:23
5 1026 3
QðzÞ
z 2 2:895z2 1 2:791z 2 0:8959
Thus, we have the polynomials
PðzÞ 5 1:622 3 1026 z2 1 45:14 3 1026 z 1 48:23 3 1026
with coefficients
p2 5 1:622 3 1026 ;
p1 5 45:14 3 1026 ;
p0 5 48:23 3 1026
QðzÞ 5 z3 2 2:895z2 1 2:791z 2 0:8959
with coefficients
q2 5 22:895;
q1 5 2:791;
q0 5 20:8959
391
CHAPTER 9 State Feedback Control
The plant is third order—that is, n 5 3—and the solvability condition of the
Diophantine equation is m 5 n 2 1 5 2. The order of the desired closed-loop characteristic
polynomial is m 1 n 5 5. We can therefore select the controller poles as {0.6, 0.4 6 j0.33}
and the observer poles as {0.1, 0.2} with the corresponding polynomials
∆dc ðzÞ 5 z3 2 1:6z2 1 0:9489z 2 0:18756
∆do ðzÞ 5 z2 2 0:3z 1 0:02
∆dcl ðzÞ 5 ∆dc ðzÞ∆do ðzÞ 5 z5 2 1:9z4 1 1:4489z3 2 0:50423z2 1 0:075246z 2 0:0037512
In other words, δ4 5 21.9, δ3 5 1.4489, δ2 5 20.50423, δ1 5 0.075246, and δ0 5 20.0037512.
Using the matrix equation (9.71) gives
3
2
1
0
0
0
0
0
7
6 22:8949
26
1
0
1:622 3 10
0
0
7
6
7
6
26
26
7
6 2:790752
22:8949
1
45:14
3
10
1:622
3
10
0
7
6
6
26
26
26 7
7
6 20:895852
2:790752
22:8949
48:23
3
10
45:14
3
10
1:622
3
10
7
6
6
26
26 7
0
20:895852
2:790752
0
48:23 3 10
45:14 3 10 5
4
0
3
2
0
2
3
20:895852
0
48:23 3 1026
0
1
d2
6 d 7 6 21:9 7
7
6 17 6
7
6 7 6
6 d0 7 6 1:4489 7
7
6 7 6
7
6 756
6 s2 7 6 20:50423 7
7
6 7 6
7
6 7 6
4 s1 5 4 0:075246 5
20:00375
s0
1.4
System: syscl
Peak amplitude: 1.06
Overshoot (%): 6.45
At time (sec): 0.08
1.2
System: syscl
Settling time (sec): 0.112
1
Amplitude
392
0.8
0.6
0.4
0.2
0
0
0.05
0.1
0.15
Time s
FIGURE 9.19
Step response for Example 9.15.
0.2
0.25
Problems
The MATLAB command linsolve gives the solution
d2 5 1; d1 5 0:9645; d0 5 0:6526; s2 5 1:8735 U 104 ; s1 5 22:9556 U 104 ; s0 5 1:2044 U 104
and the polynomials
DðzÞ 5 z2 1 0:9645z 1 0:6526
SðzÞ 5 1:8735 U 104 z2 2 2:9556 U 104 z 1 1:2044 U 104
Then, from (9.72), we have kff 5 1698.489 and the numerator polynomial
NðzÞ 5 1698:489ðz2 2 0:3z 1 0:02Þ
The step response of the control system of Figure 9.19 has a settling time of 0.1 s and
a percentage overshoot of less than 7%. The response meets all the design specifications.
Resources
D’Azzo, J.J., Houpis, C.H., 1988. Linear Control System Analysis and Design. McGraw-Hill,
New York.
Bass, R.W., Gura, I., 1983. High-order system design via statespace considerations.
Proceedings of JACC. Troy, New York (pp. 1193).
Chen, C.T., 1984. Linear System Theory and Design. HRW, New York.
Delchamps, D.F., 1988. StateSpace and InputOutput Linear Systems. Springer-Verlag.
Kailath, T., 1980. Linear Systems. Prentice Hall, Englewood Cliffs, NJ.
Kautsky, J.N., Nichols, K., Van Dooren, P., 1985. Robust pole assignment in linear state
feedback. Int. J. Control 41 (5), 11291155.
Mayne, D.Q., Murdoch, P., 1970. Modal control of linear time invariant systems. Int. J.
Control 11 (2), 223227.
Patel, R.V., Munro, N., 1982. Multivariable System Theory and Design. Pergammon Press,
Oxford, England.
PROBLEMS
9.1
Show that the closed-loop quadruple for (A, B, C, D) with the state feedback u(k) 5 2Kx(k) 1 v(k) is (A 2 BK, B, C 2 DK, D).
9.2
Show that for the pair (A, B) with state feedback gain matrix K to have the
closed-loop state matrix Acl 5 A 2 BK, a necessary condition is that for any
vector wT satisfying wTB 5 0T and wTA 5 λwT, Acl must satisfy wTAcl 5 λwT.
Explain the significance of this necessary condition. (Note that the condition is
also sufficient.)
9.3
Show that for the pair (A, B) with m 3 n state feedback gain matrix K to
have the closed-loop state matrix Acl 5 A 2 BK, a sufficient condition is
rankfBg 5 rankf½A 2 Acl jBg 5 m
Is the matrix K unique for given matrices A and Acl? Explain.
393
394
CHAPTER 9 State Feedback Control
9.4
Using the results of Problem 9.3, determine if the
be obtained using state feedback for the pair
2
3
2
1:0
0:1
0
0
A54 0
1
0:01 5 B 5 4 1
0
20:1 0:9
1
2
3
1:0
0:1
0
(a) Acl 5 4 212:2 21:2 0 5
0:01
0:01 0
2
3
0
0:1
0
(b) Acl 5 4 212:2 21:2 0 5
0:01
0:01 0
closed-loop matrix can
3
0
05
1
9.5
Show that for the pair (A, C) with observer gain matrix L to have the observer
matrix Ao 5 A 2 LC, a necessary condition is that for any vector v satisfying
Cv 5 0 and Av 5 λv, Ao must satisfy Aov 5 λv. Explain the significance of
this necessary condition. (Note that the condition is also sufficient.)
9.6
Show that for the pair (A, C) with n 3 l observer gain matrix L to have the
observer matrix Ao 5 A 2 LC, a sufficient condition is
82
39
C
<
=
rankfCg 5 rank 4 5 5 l
:
;
A 2 Ao
Is the matrix L unique for given matrices A and Ao? Explain.
9.7
Design a state feedback control law to assign the eigenvalues to the set {0,
0.1, 0.2} for the systems with
2
3
2
3
0:1 0:5 0
0:01
(a) A 5 4 2
0 0:2 5 b 5 4 0 5
0:2 1 0:4
0:005
2
3
2
3
20:2 20:2
0:4
0:01
(b) A 5 4 0:5
0
1 5 b54 0 5
0
20:4 20:4
0
9.8
Show that the eigenvalues {λi, i 5 1, . . . , n} of the closed-loop system for
the pair (A, B) with state feedback are scaled by α if we use the state feedback gains from the pole placement with the pair (A/α, B/α,). Verify this
result using MATLAB for the pair
2
3
2
3
0
1
0
0 1
A54
0
0
1 5B 5 4 0 1 5
20:005 20:11 20:7
1 1
and the eigenvalues {0.1, 0.2, 0.3} with α 5 0.5.
Problems
9.9
9.10
Using eigenvalues that are two to three times as fast as those of the plant,
design a state estimator for the system
2
3
0:2 0:3 0:2
(a) A 5 4 0
0 0:3 5 C 5 1 1 0
0:3 0 0:3
2
3
0:2 0:3 0:2
(b) A 5 4 0
0 0:3 5 C 5 1 0 0
0:3 0 0:3
Consider the system
2
0
A54
0
20:005
C 5 0:5
1
1
0
20:11
3
0
1 5
20:7
2 3
0
B5405
1
0 d50
(a) Design a controller that assigns the eigenvalues { 2 0.8, 2 0.3 6 j0.3}.
Why is the controller guaranteed to exist?
(b) Why can we design an observer for the system with the eigenvalues
{20.5, 20.1 6 j0.1}? Explain why the value (2 0.5) must be assigned.
(Hint: (s 1 0.1)2 (s 1 0.5) 5 s3 1 0.7 s2 1 0.11 s 1 0.005.)
(c) Obtain a similar system with a second-order observable subsystem, for
which an observer can be easily designed, as in Section 8.3.3. Design
an observer for the transformed system with two eigenvalues shifted
as in (b), and check your design using the MATLAB command place
or acker. Use the result to obtain the observer for the original system.
(Hint: Obtain an observer gain lr for the similar third-order system from
your design by setting the first element equal to zero. Then obtain the
observer gain for the original system using l 5 Tr lr, where Tr is the
similarity transformation matrix.)
(d) Design an observer-based feedback controller for the system with the controller and observer eigenvalues selected as in (a) and (b), respectively.
9.11
Design a reduced-order estimator state feedback controller for the discretized
system
2
0:1
A54 0
0:2
3
0 0:1
0:5 0:2 5
0 0:4
2
3
0:01
b54 0 5
0:005
to obtain the eigenvalues {0.1, 0.4 6 j0.4}.
cT 5 ½1; 1; 0
395
396
CHAPTER 9 State Feedback Control
9.12
Consider the following model of an armature-controlled DC motor, which
is slightly different from that described in Example 7.7:
2
3 2
0
x_1
4 x_2 5 5 4 0
0
x_3
32 3 2 3
0
0
x1
1 54 x2 5 1 4 0 5u
x3
211:1
10
2 3
x1
0 4 x2 5
x3
1
0
211
y5½1 0
For digital control with T 5 0.02, apply the state feedback controllers
determined in Example 9.4 in order to verify their robustness.
9.13
Consider the following model of a DC motor speed control system, which
is slightly different from that described in Example 6.9:
GðsÞ 5
1
ð1:2s 1 1Þðs 1 10Þ
For a sampling period T 5 0.02, obtain a statespace representation corresponding to the discrete-time system with DAC and ADC, then use it to
verify the robustness of the state controller described in Example 9.6.
9.14
9.15
9.16
Verify the robustness of the state controller determined in Example 9.7 by
applying it to the model shown in Problem 9.13.
~ BÞ
~ of (9.31) for the system with integral control is
Prove that the pair ðA;
controllable if and only if the pair (A, B) is controllable.
Consider the DC motor position control system described in Example 3.6,
where the (type 1) analog plant has the transfer function
GðsÞ 5
1
sðs 1 1Þðs 1 10Þ
For the digital control system with T 5 0.02, design a state feedback
controller to obtain a step response with zero steady-state error, zero
overshoot, and a settling time of less than 0.5 s.
9.17
Design a digital state feedback controller for the analog system
GðsÞ 5
2s 1 1
ð5s 1 1Þð10s 1 1Þ
with T 5 0.1 to place the closed-loop poles at {0.4, 0.6}. Show that the
zero of the closed-loop system is the same as the zero of the open-loop
system.
Computer Exercises
9.18
Write the closed-loop system statespace equations of a full-order observer
state feedback system with integral action.
9.19
Consider the continuous-time model of the overhead crane proposed in
Problem 7.10 with mc 5 1000 kg, ml 5 1500 kg, and l 5 8 m. Design a
discrete full-order observer state feedback control to provide motion of the
load without sway.
9.20
Consider the continuous-time model of the overhead crane proposed in
Problem 7.10 with mc 5 1000 kg, ml 5 1500 kg, and l 5 8 m. Design a
control system based on pole assignment using transfer functions in order
to provide motion of the load without sway.
COMPUTER EXERCISES
9.21
Write a MATLAB script to evaluate the feedback gains using Ackermann’s
formula for any pair (A, B) and any desired poles {λ1, . . . , λn}.
9.22
Write a MATLAB program that, for a multi-input system with known state
and input matrices, determines (i) a set of feasible eigenvectors and (ii) a
state feedback matrix that assign a given set of eigenvalues. Test your
program using Example 9.5.
9.23
Write a MATLAB function that, given the system statespace matrices A, B,
and C, the desired closed-loop poles, and the observer poles, determines the
closed-loop system statespace matrices of a full-observer state feedback
system with integral action.
9.24
Write a MATLAB function that uses the transfer function approach to
determine the closed-loop system transfer function for a given plant transfer
function G(z), desired closed-loop system poles, and observer poles.
397
CHAPTER
Optimal Control
10
OBJECTIVES
After completing this chapter, the reader will be able to do the following:
1. Find the unconstrained optimum values of a function by minimization or maximization.
2. Find the constrained optimum values of a function by minimization or maximization.
3. Design an optimal digital control system.
4. Design a digital linear quadratic regulator.
5. Design a digital steady-state regulator.
6. Design a digital output regulator.
7. Design a regulator to track a nonzero constant input.
8. Use the Hamiltonian system to obtain a solution of the Riccati equation.
9. Characterize the eigenstructure of the system with optimal control in terms of the
eigenstructure of the Hamiltonian system.
In this chapter, we introduce optimal control theory for discrete-time systems. We
begin with unconstrained optimization of a cost function and then generalize to
optimization with equality constraints. We then cover the optimization or optimal
control of discrete-time systems. We specialize to the linear quadratic regulator
and obtain the optimality conditions for a finite and for an infinite planning horizon. We also address the regulator problem where the system is required to track
a nonzero constant signal. Finally, we examine the Hamiltonian system, its use in
solving the Riccati equation, and its relation to the eigenstructure of the system
with optimal control.
10.1 Optimization
Many problems in engineering can be solved by minimizing a measure of cost
or maximizing a measure of performance. The designer must select a suitable
performance measure based on his or her understanding of the problem to include
the most important performance criteria and reflect their relative importance.
The designer must also select a mathematical form of the function that makes
solving the optimization problem tractable.
We observe that any maximization problem can be recast as a minimization,
or vice versa. This is because the location of the maximum of a function f(x) is
399
400
CHAPTER 10 Optimal Control
40
30
–f (x)
20
10
0
f (x)
–10
–20
–30
–40
–6
–4
–2
0
2
4
6
8 x
FIGURE 10.1
Minimization and maximization of a function of a single variable.
the same as that of the minimum 2 f(x), as demonstrated in Figure 10.1 for a
scalar x. We therefore consider minimization only throughout this chapter. We
first consider the problem of minimizing a cost function or performance measure,
then we extend our solution to problems with equality constraints.
10.1.1 Unconstrained optimization
Consider the problem of minimizing a cost function or performance measure of
the form
J ðxÞ
(10.1)
where x is an n 3 1 vector of parameters to be selected. Let the optimal parameter
vector be x and expand the function J(x) in the vicinity of the optimum as
3T
2 3 #
@J
1
@2 J JðxÞ 5 Jðx Þ 1 5 ∆x 1 ∆xT 4 2 5 ∆x 1 O jj ∆ xjj3
@x
2!
@x x
x
(10.2)
∆x 5 x 2 x
where
@J
@x
T
x
"
@J
5 @x
1
@J
@x2
@2 J @2 J 5
@x2 x
@xi @xj x
?
@J
@xn
#
(10.3)
x
(10.4)
and the subscript denotes that the matrix is evaluated at x . At a local minimum
or local maximum, the first-order terms of the expansion that appear in the gradient vector are zero.
10.1 Optimization
To guarantee that the point x is a minimum, any perturbation vector ∆x
away from x must result in an increase of the value of J(x). Thus, the secondorder term of the expansion must at least be positive or zero for x to have any
chance of being a minimum. If the second-order term is positive, then we can
guarantee that x is indeed a minimum. The sign of the second term is determined
by the characteristics of the second derivative matrix, or Hessian. The second
term is positive for any perturbation if the Hessian matrix is positive definite, and
it is positive or zero if the Hessian matrix is positive semidefinite. We summarize
this discussion in Theorem 10.1.
THEOREM 10.1
If x is a local minimum of J(x), then
@J T
5 01 3 n
@x x
@2 J $0
@x2 (10.5)
(10.6)
x
A sufficient condition for x to be a minimum is
@2 J .0
@x2 x
(10.7)
EXAMPLE 10.1
Obtain the least-squares estimates of the linear resistance
v 5 iR
using N noisy measurements
zðkÞ 5 iðkÞR 1 vðkÞ;
k 5 1; . . . ; N
Solution
We begin by stacking the measurements to obtain the matrix equation
z 5 iR 1 v
in terms of the vectors
T
z 5 zð1Þ ? zðNÞ
T
i 5 ið1Þ ? iðNÞ
T
v 5 vð1Þ ? vðNÞ
We minimize the sum of the squares of the errors
J 5
N
X
e2 ðkÞ
k51
eðkÞ 5 zðkÞ 2 iðkÞR^
where the caret (^) denotes the estimate. We rewrite the performance measure in terms of
the vectors as
2
J 5 eT e 5 zT z 2 2iT zR^ 1 iT iR^
T
e 5 eð1Þ ? eðNÞ
401
402
CHAPTER 10 Optimal Control
The necessary condition for a minimum gives
@J
5 2 2iT z 1 2iT iR^ 5 0
@R^
We now have the least-squares estimate
iT z
R^LS 5 T
i i
The solution is indeed a minimum because the second derivative is the positive sum of the
squares of the currents
@2 J
5 2iT i . 0
2
@R^
10.1.2 Constrained optimization
In most practical applications, the entries of the parameter vector x are subject to
physical and economic constraints. Assume that our vector of parameters is subject to the equality constraint
mðxÞ 5 0m 3 1
(10.8)
In the simplest cases only, we can use the constraints to eliminate m parameters
and then solve an optimization problem for the remaining n 2 m parameters.
Alternatively, we can use Lagrange multipliers to include the constraints in the
optimization problem. We add the constraint to the performance measure weighted
by the Lagrange multipliers to obtain the Lagrangian
LðxÞ 5 JðxÞ 1 λT mðxÞ
(10.9)
We then solve for the vectors x and λ that minimize the Lagrangian as in unconstrained optimization. Example 10.2 demonstrates the use of Lagrange multipliers
in constrained optimization. Note that the example is simplified to allow us to solve
the problem with and without Lagrange multipliers.
EXAMPLE 10.2
A manufacturer decides the production level of two products based on maximizing profit
subject to constraints on production. The manufacturer estimates profit using the simplified measure
JðxÞ 5 xα1 xβ2
where xi is the quantity produced for product i, i 5 1, 2, and the parameters (α, β) are
determined from sales data. The sum of the quantities of the two products produced cannot exceed a fixed level b. Determine the optimum production level for the two products
subject to the production constraint
x1 1 x2 5 b
10.1 Optimization
Solution
To convert the maximization problem into minimization, we use the negative of the profit.
We obtain the optimum first without the use of Lagrange multipliers. We solve for x2 using
the constraint
x2 5 b 2 x1
then substitute in the negative of the profit function to obtain
Jðx1 Þ 5 2 xα1 ðb2x1 Þβ
The necessary condition for a minimum gives
@J
5 2 αxα21
ðb2x1 Þβ 1 βxα1 ðb2x1 Þβ21 5 0
1
@x1
which simplifies to
x1 5
αb
α1β
From the production constraint we solve for production level
x2 5
βb
α1β
We now show that the same answer can be obtained using a Lagrange multiplier. We
add the constraint multiplied by the Lagrange multiplier to obtain the Lagrangian
LðxÞ 5 2 xα1 xβ2 1 λðx1 1 x2 2 bÞ
The necessary conditions for the minimum are
@L
5 2 αxα21
xβ2 1 λ 5 0
1
@x1
@L
5 2 βxα1 xβ21
1λ50
2
@x2
@L
5 x1 1 x2 2 b 5 0
@λ
The first two conditions give
xβ2 5 βxα1 xβ21
λ 5 αxα21
1
2
which readily simplifies to
x2 5
β
x1
α
Substituting in the third necessary condition (i.e., in the constraint) gives the solution
x1 5
αb
α1β
x2 5
βb
α1β
403
404
CHAPTER 10 Optimal Control
10.2 Optimal control
To optimize the performance of a discrete-time dynamic system, we minimize the
performance measure
J 5 Jf ðxðkf Þ; kf Þ 1
kX
f 21
LðxðkÞ; uðkÞ; kÞ
(10.10)
k5k0
subject to the constraint
xðk 1 1Þ 5 AxðkÞ 1 BuðkÞ;
k 5 k0 ; . . .; kf 2 1
(10.11)
We assume that the pair (A, B) is stabilizable; otherwise, there is no point in
control system design, optimal or otherwise. If the system is not stabilizable, then
its structure must first be changed by selecting different control variables that
allow its stabilization.
The first term of the performance measure is a terminal penalty, and each of
the remaining terms represents a cost or penalty at time k. We change the problem
into unconstrained minimization using Lagrange multipliers. We have the new
performance measure
J 5 Jf ðxðkf Þ; kf Þ 1
kX
f 21
LðxðkÞ; uðkÞ; kÞ 1 λT ðk 1 1Þ½AxðkÞ 1 BuðkÞ 2 xðk 1 1Þ
k5k0
(10.12)
We define the Hamiltonian function as
HðxðkÞ; uðkÞ; kÞ 5 LðxðkÞ; uðkÞ; kÞ 1 λT ðk 1 1Þ½AxðkÞ 1 BuðkÞ;
k 5 k0 ; . . .; kf 2 1
(10.13)
and rewrite the performance measure in terms of the Hamiltonian as
J 5 Jf ðxðkf Þ; kf Þ 2 λT ðkf Þxðkf Þ 1
kX
f 21
HðxðkÞ; uðkÞ; kÞ 2 λT ðkÞxðkÞ 1 Hðxðk0 Þ; uðk0 Þ; k0 Þ
(10.14)
k5k0 11
Each term of the preceding expression can be expanded as a truncated Taylor
series in the vicinity of the optimum point of the form
J 5 J 1 δ1 J 1 δ2 J 1 . . .
(10.15)
where the denotes the optimum, δ denotes a variation from the optimal, and the
subscripts denote the order of the variation in the expansion. From basic calculus,
we know the necessary condition for a minimum or maximum is that the firstorder term must be zero. For a minimum, the second-order term must be positive
or zero, and a sufficient condition for a minimum is a positive second term. We
therefore need to evaluate the terms of the expansion to determine necessary and
sufficient conditions for a minimum.
10.2 Optimal control
Each term of the expansion includes derivatives with respect to the arguments
of the performance measure and can be obtained by expanding each term separately. For example, the Hamiltonian can be expanded as
HðxðkÞ; uðkÞ; kÞ
T
T
@HðxðkÞ; uðkÞ; kÞ @HðxðkÞ; uðkÞ; kÞ 5 Hðx ðkÞ; u ðkÞ; kÞ 1
δxðkÞ 1
δuðkÞ
@xðkÞ
@uðkÞ
T
1½AxðkÞ1BuðkÞ δλðk 1 1Þ @2 HðxðkÞ; uðkÞ; kÞ @2 HðxðkÞ; uðkÞ; kÞ T
T
1δx ðkÞ
δxðkÞ 1 δu ðkÞ
δuðkÞ
@x2 ðkÞ
@u2 ðkÞ
@2 HðxðkÞ; uðkÞ; kÞ @2 HðxðkÞ; uðkÞ; kÞ 1δxT ðkÞ
δuðkÞ 1 δuT ðkÞ
δxðkÞ
@xðkÞ@uðkÞ
@uðkÞ@xðkÞ (10.16)
1higher-order terms
where δ denotes a variation from the optimal, the superscript ( ) denotes the optimum value, and the subscript ( ) denotes evaluation at the optimum point. The
second-order terms can be written more compactly in terms of the Hessian matrix
of second derivatives in the form
2 2
3
@ HðxðkÞ; uðkÞ; kÞ @2 HðxðkÞ; uðkÞ; kÞ
6
@x2 ðkÞ
@xðkÞ@uðkÞ 7
6
7 δxðkÞ
T
T
6
7
(10.17)
δx ðkÞ δu ðkÞ 6 2
2
7
4 @ HðxðkÞ; uðkÞ; kÞ @ HðxðkÞ; uðkÞ; kÞ 5 δuðkÞ
@uðkÞ@xðkÞ
@u2 ðkÞ
We expand the linear terms of the performance measure as
λT ðkÞxðkÞ 5 λ T ðkÞx ðkÞ 1 δλT ðkÞx ðkÞ 1 λ T ðkÞδxðkÞ
(10.18)
The terminal penalty can be expanded as
T
@J
ðxðk
Þ;
k
Þ
f
f
f
Jf ðxðkf Þ; kf Þ 5 Jf ðx ðkf Þ; kf Þ 1
δxðkf Þ 1
@xðkf Þ
2
3
@2 Jf ðx ðkf Þ; kf Þ5
δxðkf Þ
δxT ðkf Þ4
@x2 ðkf Þ
We now combine the first-order terms to obtain the first variation
82
3T
N 21 <
X
@HðxðkÞ;
uðkÞ;
kÞ
4
δ1 J 5
2λ ðkÞ5 δxðkÞ
:
@xðkÞ
k5k0 11
1 ½Ax ðkÞ1B uðkÞ2x ðk11ÞT δλðk 1 1Þ
(10.19)
405
406
CHAPTER 10 Optimal Control
2
3
T
T
@HðxðkÞ; uðkÞ; kÞ @J
ðxðk
Þ;
k
Þ
f
f
f
4
5
2λðkf Þ δxðkf Þ
1
δuðkÞ 1
@uðkÞ
@xðkf Þ
(10.20)
From (10.15) and the discussion following it, we know that a necessary condition for a minimum of the performance measure is that the first variation must be
equal to zero for any combination of variations in its arguments. A zero for any
combination of variations occurs only if the coefficient of each increment is equal
to zero. Equating each coefficient to zero, we have the necessary conditions
x ðk 1 1Þ 5 Ax ðkÞ 1 B uðkÞ;
@HðxðkÞ;
uðkÞ;
kÞ
λ ðkÞ 5
@xðkÞ
k 5 k0 ; . . .; kf 2 1
k 5 k0 ; . . .; kf 2 1
@HðxðkÞ; uðkÞ; kÞ 50
@uðkÞ
T
@Jf ðxðkf Þ; kf Þ
2λðkf Þ δxðkf Þ
05
@xðkf Þ
(10.21)
(10.22)
(10.23)
(10.24)
If the terminal point is fixed, then its perturbation is zero and the terminal
optimality condition is satisfied. Otherwise, we have the terminal conditions
λðkf Þ 5
@Jf ðxðkf Þ; kf Þ
@xðkf Þ
(10.25)
Similarly, conditions can be developed for the case of a free initial state with
an initial cost added to the performance measure.
The necessary conditions for a minimum of equations (10.21) through (10.23)
are known, respectively, as the state equation, the costate equation, and the
minimum principle of Pontryagin. The minimum principle tells us that the cost
is minimized by choosing the control that minimizes the Hamiltonian. This condition can be generalized to problems where the control is constrained, in which
case the solution can be at the boundary of the allowable control region. We summarize the necessary conditions for an optimum in Table 10.1.
Table 10.1 Optimality Conditions
Condition
Equation
State equation
k 5 k0 ; . . . ; kf 2 1
x ðk 1 1Þ 5 Ax ðkÞ 1 B uðkÞ
@HðxðkÞ;
uðkÞ;
kÞ
k 5 k0 ; . . .; kf 2 1
λ ðkÞ 5
@xðkÞ
@HðxðkÞ; uðkÞ; kÞ k 5 k0 ; . . .; kf 2 1
5 0;
@uðkÞ
Costate equation
Minimum principle
10.2 Optimal control
The second-order term is
2
3
@HðxðkÞ; uðkÞ; kÞ @HðxðkÞ; uðkÞ; kÞ
6
kX
f 21
@x2 ðkÞ
@xðkÞ@uðkÞ 7
6
7 δxðkÞ
T
T
6
7
δ2 J 5
δx ðkÞ δu ðkÞ 6 2
2
7
4 @ HðxðkÞ; uðkÞ; kÞ @ HðxðkÞ; uðkÞ; kÞ 5 δuðkÞ
k5k0
@uðkÞ@xðkÞ
@u2 ðkÞ
2
@
J
ðxðk
Þ;
k
Þ
f
f
f 1 δxT ðkf Þ
δxðkf Þ
@x2 ðkf Þ (10.26)
For the second-order term to be positive or at least zero for any perturbation,
we need the Hessian matrix to be positive definite, or at least positive semidefinite. We have the sufficient condition
2
@HðxðkÞ; uðkÞ; kÞ
6
@x2 ðkÞ
6
6 2
6 @ HðxðkÞ; uðkÞ; kÞ
4
@uðkÞ@xðkÞ
3
@HðxðkÞ; uðkÞ; kÞ
@xðkÞ@uðkÞ 7
7
7 .0
2
@ HðxðkÞ; uðkÞ; kÞ 7
5
@u2 ðkÞ
(10.27)
If the Hessian matrix is positive semidefinite, then the condition is necessary
for a minimum but not sufficient because there will be perturbations for which
the second-order terms are zero. Higher-order terms are then needed to determine
if these perturbations result in an increase in the performance measure.
For a fixed terminal state, the corresponding perturbations are zero and no
additional conditions are required. For a free terminal state, we have the additional condition
@2 Jðxðkf Þ; kf Þ (10.28)
.0
@x2 ðkf Þ EXAMPLE 10.3
Because of their preference for simple models, engineers often use first-order systems for
design. These are known as scalar systems. We consider the scalar system
xðk 1 1Þ 5 axðkÞ 1 buðkÞ; xð0Þ 5 x0 ; b . 0
If the system is required to reach the zero state, find the control that minimizes the
control effort to reach the desired final state.
Solution
For minimum control effort, we have the performance measure
f
1X
uðkÞ2
2 k50
k
J5
407
408
CHAPTER 10 Optimal Control
The Hamiltonian is given by
1
HðxðkÞ; uðkÞ; kÞ 5 uðkÞ2 1 λðk 1 1Þ½axðkÞ 1 buðkÞ;
2
k 5 0; . . . ; kf 2 1
We minimize the Hamiltonian by selecting the control
u ðkÞ 5 2 bλ ðk 1 1Þ
The costate equation is
λ ðkÞ 5 aλ ðk 1 1Þ
k 5 0; . . . ; kf 2 1
and its solution is given by
λ ðkÞ 5 akf 2k λ ðkf Þ;
k 5 0; . . . ; kf 2 1
The optimal control can now be written as
u ðkÞ 5 2 bakf 2k21 λ ðkf Þ
k 5 0; . . . ; kf 2 1
We next consider iterations of the state equation with the optimal control
x ð1Þ 5 axð0Þ 1 bu ð0Þ 5 ax0 2 b2 akf 21 λ ðkf Þ
x ð2Þ 5 axð1Þ 1 bu ð1Þ 5 a2 x0 2 b2 fakf 1 akf 22 gλ ðkf Þ
x ð3Þ 5 axð2Þ 1 bu ð2Þ 5 a3 x0 2 b2 fakf 11 1 akf 21 1 akf 23 gλ ðkf Þ
^
x ðkf Þ 5 axðkf 2 1Þ 1 bu ðkf 2 1Þ 5 akf x0 2 b2 fða2 Þkf 21 1 . . . 1a2 1 1gλ ðkf Þ 5 0
We solve the last equation for the terminal Lagrange multiplier
ðakf =b2 Þx0
ða2 Þkf 21 1 . . . 1 a2 1 1
λ ðkf Þ 5
Substituting for the terminal Lagrange multiplier gives the optimal control
u ðkÞ 5 2
ða2kf 2k21 =bÞx0
ða2 Þkf 21 1 . . . 1 a2 1 1
k 5 0; . . . ; kf 2 1
For an open-loop stable system, we can write the control in the form
u ðkÞ 5 2
a2 2 1
a2kf 2k21 =b x0
ða2 Þkf 2 1
k 5 0; . . . ; kf 2 1; jaj , 1
The closed-loop system is given by
xðk 1 1Þ 5 axðkÞ 2
a2kf 2k21 =b x0
k 21
a2 f
1 . . . 1 a2 1 1
;
xð0Þ 5 x0 ; b . 0
Note that the closed-loop system dynamics of the optimal system are time varying
even though the open-loop system is time invariant.
10.3 The linear quadratic regulator
x (k)
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0
1
2
3
4
5
6
7
8
9
10 k
FIGURE 10.2
Optimal trajectory for the scalar system of Example 10.3.
For an initial state x0 5 1 and the unstable dynamics with a 5 1.5 and b 5 0.5, the optimal trajectories are shown in Figure 10.2. The optimal control stabilizes the system and
drives its trajectories to the origin.
10.3 The linear quadratic regulator
The choice of performance index in optimal control determines the performance
of the system and the complexity of the optimal control problem. The most popular choice for the performance index is a quadratic function of the state variable
and the control inputs. The performance measure is of the form
k 21
J5
f
1 T
1X
xT ðkÞQðkÞxðkÞ 1 uT ðkÞRðkÞuðkÞ
x ðkf ÞSðkf Þxðkf Þ 1
2
2 k5k
(10.29)
0
where the matrices S(kf), Q(k), k 5 k0, . . . , kf 2 1 are positive semidefinite symmetric n 3 n, and the matrices R(k), k 5 k0, . . . , kf 2 1 are positive definite symmetric m 3 m. This choice is known as the linear quadratic regulator, and in
addition to its desirable mathematical properties, it can be physically justified.
In a regulator problem, the purpose is to maintain the system close to the zero
state, and the quadratic function of the state variable is a measure of error. On the
other hand, the effort needed to minimize the error must also be minimized, as
must the control effort represented by a quadratic function of the controls. Thus,
the quadratic performance measure achieves a compromise between minimizing
the regulator error and minimizing the control effort.
409
410
CHAPTER 10 Optimal Control
To obtain the necessary and sufficient conditions for the linear quadratic regulator, we use the results of Section 10.2. We first write an expression for the
Hamiltonian:
HðxðkÞ; uðkÞ; kÞ 5
1 T
1
x ðkÞQðkÞxðkÞ 1 uT ðkÞRðkÞuðkÞ
2
2
1 λ ðk 1 1Þ½AxðkÞ 1 BuðkÞ;
T
(10.30)
k 5 k0 ; . . .; kf 2 1
The state equation is unchanged, but the costate equation becomes
λ ðkÞ 5 QðkÞx ðkÞ 1 AT λ ðk 1 1Þ
k 5 k0 ; . . .; kf 2 1
(10.31)
Pontryagin’s minimum principle gives
RðkÞu ðkÞ 1 BT λ ðk 1 1Þ 5 0
which yields the optimum control expression
u ðkÞ 5 2 R21 ðkÞBT λ ðk 1 1Þ
(10.32)
Substituting in the state equation, we obtain the optimal dynamics
x ðk 1 1Þ 5 Ax ðkÞ 2 B R21 ðkÞBT λ ðk 1 1Þ;
Equation (10.27) gives the condition
QðkÞ
0m 3 n
k 5 k0 ; . . .; kf 2 1
0n 3 m
.0
R
(10.33)
(10.34)
Because R is positive definite, a sufficient condition for a minimum is that Q
must be positive definite. A necessary condition is that Q must be positive
semidefinite.
10.3.1 Free final state
If the final state of the optimal control is not fixed, we have the terminal condition based on (10.25):
λ ðkf Þ 5
@Jf ðxðkf Þ; kf Þ
5 Sðkf Þx ðkf Þ
@xðkf Þ
(10.35)
where the matrix S(kf) is the terminal weight matrix of (10.29). The condition
suggests a general relationship between the state and costate—that is,
λ ðkÞ 5 SðkÞx ðkÞ
(10.36)
If the proposed relation yields the correct solution, this would allow us to conclude that the relation is indeed correct. We substitute the proposed relation
(10.36) in the state equation (10.33):
x ðk 1 1Þ 5 Ax ðkÞ 2 BR21 ðkÞBT Sðk 1 1Þx ðk 1 1Þ
10.3 The linear quadratic regulator
This yields the recursion
x ðk 1 1Þ 5 fI1BR21 ðkÞBT Sðk11Þg21 Ax ðkÞ
(10.37)
Using the proposed relation (10.36) and substituting in the costate equation,
we have
QðkÞx ðkÞ 1 AT Sðk 1 1Þx ðk 1 1Þ
λ ðkÞ 5 5 QðkÞ 1 AT Sðk 1 1Þ½I1BR21 ðkÞBT Sðk11Þ21 A x ðkÞ
5 SðkÞx ðkÞ
(10.38)
Equation (10.38) holds for all values of the state vector and hence we have
the matrix equality
QðkÞ 1 AT Sðk 1 1Þ I1B R21 ðkÞBT Sðk11Þ
21
A 5 SðkÞ
We now apply the matrix inversion lemma (see Appendix III) to obtain
SðkÞ 5 AT fSðk 1 1Þ 2 Sðk 1 1ÞBðBT Sðk 1 1ÞB 1RðkÞÞ21 BT Sðk 1 1ÞgA
1 QðkÞ; Sðkf Þ
(10.39)
The preceding equation is known as the matrix Riccati equation and can be
solved iteratively backward in time to obtain the matrices S(k), k 5 k0,. . ., kf 2 1.
We now return to the expression for the optimal control (10.32) and substitute
for the costate to obtain
u ðkÞ 5 2R21 ðkÞBT Sðk 1 1Þx ðk 1 1Þ
5 2R21 ðkÞBT Sðk 1 1Þ½Ax ðkÞ 1 BuðkÞ
We solve for the control
u ðkÞ 5 2 I1R21 ðkÞBT Sðk11ÞB
21 21
R ðkÞBT Sðk 1 1ÞAx ðkÞ
Using the rule for the inverse of a product, we have the optimal state feedback
u ðkÞ 5 2KðkÞx ðkÞ
KðkÞ 5 ½RðkÞ1BT Sðk11ÞB21 BT Sðk 1 1ÞA
(10.40)
Thus, with the offline solution of the Riccati equation, we can implement the
optimal state feedback.
The Riccati equation can be written in terms of the optimal gain of (10.40) in
the form
SðkÞ 5 AT Sðk 1 1ÞfA 2 BðBT Sðk11ÞB1RðkÞÞ21 BT Sðk 1 1ÞAg 1 QðkÞ
5 AT Sðk 1 1ÞfA 2 BKðkÞg 1 QðkÞ
In terms of the closed-loop state matrix Acl(k), we have
SðkÞ 5 AT Sðk 1 1ÞAcl ðkÞ 1 QðkÞ
Acl ðkÞ 5 A 2 BKðkÞ
(10.41)
411
412
CHAPTER 10 Optimal Control
Note that the closed-loop matrix is time varying even for a time-invariant pair
(A, B).
A more useful form of the Riccati is obtained by adding and subtracting terms
to obtain
SðkÞ 5 ½A2BKðkÞT Sðk 1 1ÞATcl ðkÞ 1 QðkÞ 1 K T ðkÞBT ðkÞSðk 1 1Þ½A 2 BKðkÞ
Using (10.40), the added term can be rewritten as
K T ðkÞBT ðkÞSðk 1 1Þ½A2BKðkÞ
5 K T ðkÞBT ðkÞSðk 1 1ÞA 2K T ðkÞBT ðkÞSðk 1 1ÞBKðkÞ
5K T ðkÞBT ðkÞSðk 11ÞA
2 K T ðkÞBT ðkÞSðk 1 1ÞB½RðkÞ1BT Sðk11ÞB21 BT Sðk 1 1ÞA
5 K T ðkÞBT ðkÞSðk 11ÞA
1K T ðkÞfRðkÞ2 ½RðkÞ 1BT ðkÞSðk 1 1ÞBg½RðkÞ1BT Sðk11ÞB21 BT Sðk 1 1ÞA
5 K T ðkÞRðkÞKðkÞ
We now have the Joseph form of the Riccati equation:
SðkÞ 5 ATcl ðkÞSðk 1 1ÞAcl ðkÞ 1 K T ðkÞRðkÞKðkÞ 1 QðkÞ
(10.42)
Because of its symmetry, this form performs better in iterative numerical
computation.
If the Riccati equation in Joseph form is rewritten with the optimal gain
replaced by a suboptimal gain, then the equation becomes a Lyapunov difference
equation (see Chapter 11). For example, if the optimal gain is replaced by a constant gain K to simplify implementation, then the performance of the system will
be suboptimal and is governed by a Lyapunov equation.
EXAMPLE 10.4
A variety of mechanical systems can be approximately modeled by the double integrator
x€ ðtÞ 5 uðtÞ
where u(t) is the applied force. With digital control and a sampling period T 5 0.02 s, the
double integrator has the discrete statespace equation
1 T
T 2 =2
uðkÞ
xðkÞ 1
xðk 1 1Þ 5
0 1
T
yðkÞ 5 1 0 xðkÞ
1. Design a linear quadratic regulator for the system with terminal weight S(100) 5
diag{10, 1}, Q 5 diag{10, 1}, and control weight r 5 0.1, then simulate the system with
initial condition vector x(0) 5 [1, 0]T.
2. Repeat the design of part 1 with S(100) 5 diag{10, 1}, Q 5 diag{10, 1}, and control
weight r 5 100. Compare the results to part 1, and discuss the effect of changing the
value of r.
10.3 The linear quadratic regulator
Solution
The Riccati equation can be expanded with the help of a symbolic manipulation package
to obtain
"
#
s11 s12
s12
"
s22
k
#
2
3
2
4
2
2ðs11 s22 2s12 ÞT 14rðs11 T 12s12 Þ ðs11 s22 2s12 ÞT 14ðs11 T 12Ts12 1s22 Þr 5
T 4 s11 14T 3 s12 14T 2 s22 14r
ðs11 s22 2s212 ÞT 2 1s11 r
2ðs11 s22 2s212 ÞT 3 14rðs11 T 12s12 Þ
k11
1Q
where the subscript denotes the time at which the matrix S is evaluated by backward-intime iteration starting with the given terminal value. We use the simple MATLAB script:
% simlqr: simulate a scalar optimal control DT system
t(1) 5 0;
x{1} 5 [1;0]; % Initial state
T 5 0.02; % Sampling period
N 5 150; % Number of steps
S 5 diag([10,1]);r 5 0.1; % Weights
Q 5 diag([10,1]);
A 5 [1,T;0,1];B 5 [T^2/2;T]; % System matrices
for i 5 N:-1:1
K{i} 5 (r 1 B’ S A B)\B’ S A; % Calculate the optimal feedback
% gains
% Note that K(1) is really K(0)
kp(i) 5 K{i}(1); % Position gain
kv(i) 5 K{i}(2); % Velocity gain
Acl 5 A-B K{i};
S 5 Acl’ S Acl 1 K{i}’ r K{i} 1 Q; % Iterate backward (Riccati% Joseph form)
end
for i 5 1:N
t(i 1 1) 5 t(i) 1 T;
u(i) 5 -K{i} x{i};
x{i 1 1} 5 A x{i} 1 B u(i); % State equation
end
xmat 5 cell2mat(x); % Change cell to mat to extract data
xx 5 xmat(1,:); % Position
xv 5 xmat(2,:); % Velocity
plot(t,xx,’.’) % Plot Position
figure % New figure
plot(t,xv,’.’) % Plot Velocity
figure % New figure
plot(xx,xv,’.’)% Plot phase plane trajectory
figure % New figure
plot(t(1:N),u,’.’) % Plot control input
figure % New figure
plot(t(1:N),kp,’.’,t(1:N),kv,’.’) % Plot position gain & velocity
%gain
413
414
CHAPTER 10 Optimal Control
x1(t)
1
0.8
0.6
0.4
0.2
0
0
0.2 0.4 0.6 0.8
1
1.2 1.4 1.6 1.8
2 t
FIGURE 10.3
Position trajectory for the inertial control system described in Example 10.4 (1).
x2(t)
0
–0.2
–0.4
–0.6
–0.8
–1
–1.2
–1.4
0
0.2 0.4 0.6 0.8
1
1.2 1.4 1.6 1.8
2 t
FIGURE 10.4
Velocity trajectory for the inertial control system described in Example 10.4 (1).
10.3 The linear quadratic regulator
x2(t)
0
–0.2
–0.4
–0.6
–0.8
–1
–1.2
–0.2
0
0.2
0.4
0.6
0.8
1
1.2 x1(t)
FIGURE 10.5
Phase plane trajectory for the inertial control system described in Example 10.4 (1).
K
9
8
Position gain K(1)
7
6
5
Velocity gain K(2)
4
3
2
1
0
0
0.2
0.4 0.6
0.8
1
1.2 1.4
1.6 1.8
2t
FIGURE 10.6
Plot of feedback gains versus time for Example 10.4 (1).
Five plots are obtained using the script for each part.
1. The first three plots are the position trajectory of Figure 10.3, the velocity trajectory of
Figure 10.4, and the phase-plane trajectory of Figure 10.5. The three plots show that
the optimal control drives the system toward the origin. However, because the final
415
416
CHAPTER 10 Optimal Control
u(t)
2
0
–2
–4
–6
–8
–10
0
0.2
0.4 0.6
0.8
1
1.2 1.4
1.6 1.8
2t
4
5 t
FIGURE 10.7
Optimal control input for Example 10.4 (1).
x1(t)
1.2
1
0.8
0.6
0.4
0.2
0
–0.2
0
0.5
1
1.5
2
2.5
3
3.5
4.5
FIGURE 10.8
Position trajectory for the inertial control system of Example 10.4 (2).
state is free, the origin is not reached. From the gain plot shown in Figure 10.6, we
observe that the controller gains, which are obtained by iteration backward in time,
approach a fixed level. Consequently, the optimal control shown in Figure 10.7 also
approaches a fixed level.
10.3 The linear quadratic regulator
x2(t)
0
–0.05
–0.1
–0.15
–0.2
–0.25
–0.3
–0.35
0
0.5
1
1.5
2
2.5
3
3.5
4
4.5
5 t
FIGURE 10.9
Velocity trajectory for the inertial control system of Example 10.4 (2).
x2(t)
0
–0.05
–0.1
–0.15
–0.2
–0.25
–0.3
–0.35
–0.2
0
0.2
0.4
0.6
0.8
1 x1(t)
FIGURE 10.10
Phase plane trajectory for the inertial control system of Example 10.4 (2).
2. The first three plots are the position trajectory of Figure 10.8, the velocity trajectory of
Figure 10.9, and the phase-plane trajectory of Figure 10.10. The three plots show that
the optimal control drives the system toward the origin. However, the final state
reached is farther from the origin than that of part 1 because the error weight matrix is
417
418
CHAPTER 10 Optimal Control
K
0.8
0.7
Position gain K(1)
0.6
0.5
0.4
0.3
0.2 Velocity gain K(2)
0.1
0
0
0.5
1
1.5
2
2.5
3
3.5
4
4.5
5 t
FIGURE 10.11
Plot of feedback gains versus time for Example 10.4 (2).
u(t)
0.05
0
–0.05
–0.1
–0.15
–0.2
–0.25
–0.3
–0.35
0
0.5
1
1.5
2
2.5
3
3.5
4
4.5
5t
FIGURE 10.12
Optimal control input for Example 10.4 (2).
now smaller relative to the value of r. The larger control weight r results in smaller control gain as shown in Figure 10.11. This corresponds to a reduction in the control effort
as shown in Figure 10.12. Note that the controller gains approach a fixed level at a
slower rate than in part 1.
10.4 Steady-state quadratic regulator
10.4 Steady-state quadratic regulator
Implementing the linear quadratic regulator is rather complicated because of the
need to calculate and store gain matrices. From Example 10.4 we observe that the
gain values converge to fixed values. This occurs in general with the fulfillment
of some simple requirements that are discussed in this section.
For many applications, it is possible to simplify implementation considerably
by using the steady-state gain exclusively in place of the optimal gains. This solution is only optimal if the summation interval in the performance measure, known
as the planning horizon, is infinite. For a finite planning horizon, the simplified
solution is suboptimal (i.e., gives a higher value of the performance measure) but
often performs almost as well as the optimal control. Thus, it is possible to retain
the performance of optimal control without the burden of implementing it if we
solve the steady-state regulator problem with the performance measure of
the form
J5
N
1X
ðxT ðkÞQxðkÞ 1 uT ðkÞRuðkÞÞ
2 k5k
(10.43)
0
We assume that the weighting matrices Q and R are constant, with Q positive
semidefinite and R positive definite. We can therefore decompose the matrix Q as
Q 5 QTa Qa
(10.44)
where Qa is the square root matrix. This allows us to write the state error terms
of “measurements” in terms of an equivalent measurement vector
yðkÞ 5 Qa xðkÞ
xT ðkÞQxðkÞ 5 yT ðkÞyðkÞ
(10.45)
The matrix Qa is positive definite for Q positive definite and positive semidefinite for Q positive semidefinite. In the latter case, large state vectors can be
mapped by Qa to zero y(k), and a small performance measure cannot guarantee
small errors or even stability. We think of the vector y(k) as a measurement of
the state and recall the detectability condition from Chapter 8.
We recall that if the pair (A, Qa) is detectable, then x(k) decays asymptotically
to zero with y(k). Hence, this detectability condition is required for
acceptable steady-state regulator behavior. If the pair is observable, then it is also
detectable and the system behavior is acceptable. For example, if the matrix Qa is
positive definite, there is a oneone mapping between the states x and the measurements y, and the pair (A, Qa) is always observable. Hence, a positive definite
Q (and Qa) is sufficient for acceptable system behavior.
If the detectability (observability) condition is satisfied and the system is stabilizable, then the Riccati equation does not diverge and a steady-state condition is
reached. The resulting algebraic Riccati equation is in the form
(10.46)
S 5 AT S 2 SBðBT SB1RÞ21 BT S A 1 Q
419
420
CHAPTER 10 Optimal Control
It can be shown that, under these conditions, the algebraic Riccati equation
has a unique positive definite solution. However, the equation is clearly difficult
to solve in general and is typically solved numerically. The MATLAB solution of
the algebraic Riccati equation is discussed in Section 10.4.2.
The optimal state feedback corresponding to the steady-state regulator is
u 5 2 Kx ðkÞ
K 5 ½R1BT SB21 BT SA
(10.47)
Using the gain expression, we can write the algebraic Riccati equation in the
Joseph form:
S 5 ATcl SAcl 1 K T RK 1 Q
Acl 5 A 2 BK
(10.48)
If the optimal gain K is replaced by a suboptimal gain, then the algebraic
Riccati equation becomes an algebraic Lyapunov equation (see Chapter 11).
The Lyapunov equation is clearly linear and its solution is simpler than that of
the Riccati equation.
10.4.1 Output quadratic regulator
In most practical applications, the designer is interested in optimally controlling
the output y(k) rather than the state x(k). To optimally control the output, we need
to consider a performance index of the form
J5
N
1X
ðyT ðkÞQy yðkÞ 1 uT ðkÞRuðkÞÞ
2 k5k
(10.49)
0
From (10.45), we observe that this is equivalent to the original performance
measure of (10.43) with the state weight matrix
Q 5 C T Qy C
5 C T QTya Qya C
5 QTa Qa
Qa5 Qya C
(10.50)
where Qya is the square root of the output weight matrix. As in the state quadratic regulator, the Riccati equation for the output regulator can be solved using
the MATLAB commands discussed in Section 10.4.2. For a stabilizable pair
(A, B), the solution exists provided that the pair (A, Qa) is detectable with Qa as
in (10.50).
10.4 Steady-state quadratic regulator
10.4.2 MATLAB solution of the steady-state regulator problem
MATLAB has several commands that allow us to conveniently design steadystate regulators. The first is dare, which solves the discrete algebraic Riccati
equation (10.40). The command is
.. ½S;E;K 5 dareðA; B; Q; RÞ
The input arguments are the state matrix A, the input matrix B, and the
weighting matrices Q and R. The output arguments are the solution of the discrete
algebraic Riccati equation S, the feedback gain matrix K, and the eigenvalues E
of the closed-loop optimal system A 2 BK.
The second command for discrete optimal control is dlqr, which solves the
steady-state regulator problem. The command has the form
.. ½K; S; E 5 dlqrðA; B; Q; RÞ
where the input arguments are the same as the command dare, and the output
arguments, also the same, are the gain K, the solution of the discrete algebraic
Riccati equation S, and the eigenvalues E of the closed-loop optimal system
A 2 BK.
For the output regulator problem, we can use the commands dare and dlqr
with the matrix Q replaced by CTQyC. Alternatively, MATLAB provides the
command
.. ½Ky; S;E 5 dlqryðA; B; C; D; Qy; RÞ
EXAMPLE 10.5
Design a steady-state regulator for the inertial system of Example 10.4, and compare its
performance to the optimal control by plotting the phase plane trajectories of the two systems. Explain the advantages and disadvantages of the two controllers.
The pair (A, B) is observable. The state-error weighting matrix is positive definite, and
the pair (A, Qa) is observable. We are therefore assured that the Riccati equation will have
a steady-state solution. For the inertial system presented in Example 10.4, we have the
MATLAB output
.. ½K;S;E 5 dlqrðA;B;Q;rÞ
K 5 9:4671 5:2817
S 5 278:9525 50:0250 50:0250 27:9089
E 5 0:9462 1 0:0299i
0:9462 2 0:0299i
If we simulate the system with the optimal control of Example 10.4 and superimpose
the trajectories for the suboptimal steady-state regulator, we obtain Figure 10.13. The
figure shows that the suboptimal control, although much simpler to implement, provides
421
422
CHAPTER 10 Optimal Control
x2(t)
0.2
0
–0.2
–0.4
–0.6
–0.8
–1
–1.2
–1.4
–0.2
0
0.2
0.4
0.6
0.8
1
1.2 x1(t)
FIGURE 10.13
Phase plane trajectories for the inertial system of Example 10.5 with optimal control (dark
gray) and suboptimal steady-state regulator (light gray).
an almost identical trajectory to that of the optimal control. For practical implementation,
the suboptimal control is often preferable because it is far cheaper to implement and provides almost the same performance as the optimal control. However, there may be situations where the higher accuracy of the optimal control justifies the additional cost of its
implementation.
EXAMPLE 10.6
Design a digital output regulator for the double integrator system
1 T
T 2 =2
uðkÞ
xðk 1 1Þ 5
xðkÞ 1
0 1
T
yðkÞ 5 1 0 xðkÞ
with sampling period T 5 0.02 s, output weight Qy 5 1, and control weight r 5 0.1, and plot
the output response for the initial condition vector x(0) 5 [1, 0]T.
Solution
We use the MATLAB command
.. ½Ky; S; E 5 dlqryðA; B; C; 0; 1; 0:1Þ
Ky 5 3:0837 2:4834
S 5 40:2667 15:8114 15:8114 12:5753
E5 0:9749 1 0:0245i 0:9749 0:0245i
10.4 Steady-state quadratic regulator
y(t)
1
0.8
0.6
0.4
0.2
0
0
0.2 0.4 0.6 0.8
1
1.2 1.4 1.6 1.8
2 t
FIGURE 10.14
Time response of the output regulator discussed in Example 10.6.
The response of the system to initial conditions x(0) 5 [1, 0]T is shown in
Figure 10.14. The output quickly converges to the target zero state.
10.4.3 Linear quadratic tracking controller
The minimization of the performance index (10.43) yields an optimal state feedback matrix K that minimizes integral square error and control effort (see
(10.47)). The error is defined relative to the zero state. In many practical applications, the system is required to follow a specified function of time. The design of
a controller to achieve this objective is known as the tracking problem. If the
control task is to track a nonzero constant reference input, we can exploit the
techniques described in Section 9.3 to solve the new optimal regulator problem.
In Section 9.3, we showed that the tracking or servo problem can be solved
using an additional gain matrix for the reference input as shown in Figure 9.7,
thus providing an additional degree of freedom in our design. For a square system
(equal number of inputs and outputs), we can implement the degree-of-freedom
scheme of Figure 9.7 with the reference gain matrix
F 5 ½CðIn 2Acl Þ21 B21
(10.51)
423
424
CHAPTER 10 Optimal Control
where
Acl 5 A 2 BK
(10.52)
All other equations for the linear quadratic regulator are unchanged.
Alternatively, to improve robustness, but at the expense of a higher-order controller, we can introduce integral control, as in Figure 9.9. In particular, as in
(9.29) we consider the statespace equations
xðk 1 1Þ 5 AxðkÞ 1 BuðkÞ
xðk 1 1Þ 5 xðkÞ 1 rðkÞ 2 yðkÞ
yðkÞ 5 CxðkÞ
uðkÞ 5 2KxðkÞ 2 KxðkÞ
(10.53)
with integral control gain K. This yields the closed-loop statespace equations
(9.30)
0
~
~
~
rðkÞ
x~ ðk 1 1Þ 5 ðA 2 BKÞ~xðkÞ 1
Il
yðkÞ 5 C 0 ~xðkÞ
where x~ ðkÞ 5 xðkÞ xðkÞ T and the matrices are given by
A
0
A~ 5
2C Il
B
~
K~ 5 K K B5
C~ 5 C 0
0
(10.54)
(10.55)
The state feedback gain matrix can be computed as
T
~ 21 B~T SA~
K~ 5 ½R1 B~ SB
(10.56)
EXAMPLE 10.7
Design an optimal linear quadratic statespace tracking controller for the inertial system
of Example 10.4 to obtain zero steady-state error due to a unit step input.
Solution
The statespace matrices are
1 0:02
0
1
C5 1 0
A5
B5
0:0002
0:02
Adding integral control, we calculate the augmented matrices
A~ 5
A
2C
2
1
0
540
1
1
3
0:02 0
1
05
0
1
2
3
0:0002
B
B~ 5
5 4 0:02 5
0
0
10.4 Steady-state quadratic regulator
We select the weight matrices with larger error weighting as
2
3
10 0 0
4
Q 5 0 1 0 5 r 5 0:1
0 0 1
The MATLAB command dlqr yields the desired solution:
.. [Ktilde, S, E] 5 dlqr(Atilde, Btilde, Q, r)
Ktilde 5 57.363418836015583 10.818259776359207
2 2.818765217602360
S 5 1.0e 1 003
2.922350387967343 0.314021626641202 2 0.191897141854919
0.314021626641202 0.058073322228013 20.015819292019556
2 0.191897141854919 2 0.015819292019556 0.020350548700473
E 5 0.939332687303670 1 0.083102739623114i
0.939332687303670 2 0.083102739623114i
0.893496746098272
The closed-loop system statespace equation is
~ xðkÞ 1
x~ ðk 1 1Þ 5 ðA~ 2 B~KÞ~
yðkÞ 5 C
0
rðkÞ
Il
0 ~xðkÞ
1
Amplitude
0.8
0.6
0.4
0.2
0
0
0.2
0.4
0.6
0.8
1
Time s
1.2
1.4
1.6
1.8
FIGURE 10.15
Step response for Example 10.7.
The closed-loop step response shown in Figure 10.15 is satisfactory with a small overshoot and zero steady-state error due to step.
425
426
CHAPTER 10 Optimal Control
10.5 Hamiltonian system
In this section, we obtain a solution of the steady-state regulator problem using
linear dynamics without directly solving the Riccati equation of (10.46). We
show that we can obtain the solution of the Riccati equation from the state-transition matrix of a linear system. We consider the case of constant weighting matrices. We can combine the state and costate equations to obtain the 2n-dimensional
Hamiltonian system:
x ðk 1 1Þ
A
5
λ ðkÞ
Q
2B R21 BT
AT
x ðkÞ
;
λ ðk 1 1Þ
k 5 k0 ; . . .; kf 2 1 (10.57)
If the state matrix A is obtained by discretizing a continuous-time system, then
it is a state transition matrix and is always invertible. We can then rewrite the
state equation in the same form as the equation of the costate l and obtain the
Hamiltonian system
x ðk 1 1Þ
x ðkÞ
5
H
;
k 5 k0 ; . . .; kf 2 1
λ ðkÞ
λ ðk 1 1Þ
A21 B R21 BT
A21
H5
QA21 AT 1 QA21 B R21 BT
λ ðkf Þ 5 Sðkf Þxðkf Þ
(10.58)
The Hamiltonian system of (10.58) describes the state and costate evolution
backward in time because it provides an expression for evaluating the vector at
time k from the vector at time k 1 1. We can solve the linear equation and write
its solution in terms of the state-transition matrix for the Hamiltonian
x ðkf Þ
x ðkÞ
kf 2k
;
k 5 k0 ; . . .; kf
5
H
λ ðkf Þ
λ ðkÞ
φ ðkf 2 kÞ φ12 ðkf 2 kÞ
H kf 2k 5 11
φ21 ðkf 2 kÞ φ22 ðkf 2 kÞ
(10.59)
The solution of the state equation yields
x ðkÞ 5 fφ11 ðkf 2 kÞ 1 φ12 ðkf 2 kÞSðkf Þgx ðkk Þ
λ ðkÞ 5 fφ21 ðkf 2 kÞ 1 φ22 ðkf 2 kÞSðkf Þgx ðkk Þ
(10.60)
We solve for the costate in terms of the state vector
λ ðkÞ 5 MðkÞx ðkÞ
MðkÞ 5 fφ21 ðkf 2 kÞ 1 φ22 ðkf 2 kÞSðkf Þgfφ11 ðkf 2 kÞ
1φ12 ðkf 2 kÞSðkf Þg21
(10.61)
10.5 Hamiltonian system
Substituting in (10.32), we now have the control
u ðkÞ 5 2R21 BT Mðk 1 1Þx ðk 1 1Þ
5 2R21 BT Mðk 1 1ÞfAx ðkÞ 1 Bu ðkÞg
k 5 k0 ; . . .; kf 2 1
(10.62)
We multiply both sides by the matrix R and then solve for the control
u ðkÞ 5 KðkÞx ðkÞ
KðkÞ 5 2 fR1BT Mðk11ÞBg21 BT Mðk 1 1ÞA
k 5 k0 ; . . .; kf 2 1
(10.63)
We compare the gain expression of (10.63) with that obtained by solving the
Riccati equation in (10.40). Because the two expressions are solutions of the
same problem and must hold for any state vector, we conclude that they must be
equal and that M(k 1 1) is identical to S(k 1 1). Hence, the solution of the Riccati
equation is given in terms of the state-transition matrix of the Hamiltonian system
by (10.58). We can now write
SðkÞ 5 fφ21 ðkf 2 kÞ 1 φ22 ðkf 2 kÞSðkf Þgfφ11 ðkf 2kÞ1φ12 ðkf 2kÞSðkf Þg21
(10.64)
It may seem that the expression of (10.64) is always preferable to direct solution of the Riccati equation because it eliminates the nonlinearity. However,
because the dimension of the Hamiltonian system is 2n, the equation doubles the
dimension of the problem. In many cases, it is preferable to solve the Riccati
equation in spite of its nonlinearity.
EXAMPLE 10.8
Consider the double integrator described in Example 10.4 with a sampling period
T 5 0.02 s, terminal weight S(10) 5 diag{10, 1}, Q 5 diag{10, 1}, and control weight r 5 0.1.
Obtain the Hamiltonian system and use it to obtain the optimal control.
Solution
The Hamiltonian system is
x ðk 1 1Þ
x ðkÞ
5
H
;
λ ðkÞ
λ ðk 1 1Þ
H5
A21
QA21
k 5 k0 ; . . .; kf 2 1
2
1
6
6 0
A21 B R21 BT
56
6
AT 1 QA21 B R21 BT
4 10
λ ð10Þ 5 diagf10; 1gxð10Þ
0
20:02
1
0
0
0:004
20:2
1
20:0004
1
0:02
1:004
0
3
7
7
7
7
5
Because the dynamics describe the evolution backward in time, each multiplication of
the current state by the matrix H yields the state at an earlier time. We define a backwardtime variable to use in computing as
kb 5 k 2ðkf 2 1Þ
427
428
CHAPTER 10 Optimal Control
Proceeding backward in time, kb starts with the value zero at k 5 kf 2 1, then increases
to kb 5 kf 2 1 with k 5 0. We solve for the transition matrices and the gains using the following MATLAB program:
% hamilton
% Form the Hamiltonian system for backward dynamics
[n,n] 5 size(a); % Order of state matrix
n2 5 2 n; % Order of Hamiltonian matrix
kf 5 11; % Final time
q 5 diag([10,1]);r 5 .1; % Weight matrices
sf 5 q; % Final state weight matrix
% Calculate the backward in time Hamiltonian matrix
a1 5 inv(a);
He 5 b/r b’;
H3 5 q/a;
H 5 [a1,a1 He;H3,a’ 1 H3 He];% Hamiltonian matrix
fi 5 H;% Initialize the state-transition matrix
% i is the backward time variable kb 5 k-(kf-1), k 5 discrete time
for i 5 1:kf-1
fi11 5 fi(1:n,1:n); % Partition the state-transition matrix
fi12 5 fi(1:n,n 1 1:n2);
fi21 5 fi(n 1 1:n2,1:n);
fi22 5 fi(n 1 1:n2,n 1 1:n2);
s 5 (fi21 1 fi22 sf)/(fi11 1 fi12 sf);% Compute the Riccati egn.
% solution
K 5 (r 1 b’ s b)\b’ s a % Calculate the gains
fi 5 H fi;% Update the state-transition matrix
end
Table 10.2 Optimal Gains for the Integrator with a Planning
Horizon kf 5 10
Time
Optimal Gain
Vector K
0
1
2
3
4
5
6
7
8
9
2.0950
1.7717
1.4656
1.1803
0.9192
0.6855
0.4823
0.3121
0.1771
0.0793
2.1507
1.9632
1.7730
1.5804
1.3861
1.1903
0.9935
0.7958
0.5976
0.3988
The optimal control gains are given in Table 10.2. Almost identical results are
obtained using the program described in Example 10.4. Small computational errors result
in small differences between the results of the two programs.
10.5 Hamiltonian system
10.5.1 Eigenstructure of the Hamiltonian matrix
The expression for the Lagrange multiplier in (10.57) can be rewritten as
λ ðk 1 1Þ 5 2A2T Qx ðkÞ 1 A2T λ ðkÞ
Substituting in the expression for the state in (10.57) allows us to write it in
terms of the variables at time k in the form
x ðk 1 1Þ 5 ½A 1 BR21 BT A2T Qx ðkÞ 2 BR21 BT A2T λ ðkÞ
We now write the forward expression for the state and Lagrange multiplier
vectors
A 1 BR21 BT A2T Q 2BR21 BT A2T
x ðkÞ
x ðk 1 1Þ
5
(10.65)
2T
2T
2A Q
A
λ ðk 1 1Þ
λ ðkÞ
Clearly, based on the backward recursion of (10.58), the inverse of the
Hamiltonian matrix is
A 1 BR21 BT A2T Q 2BR21 BT A2T
21
H 5
(10.66)
2A2T Q
A2T
In fact, the same matrix is obtained more laboriously by directly inverting the
Hamiltonian matrix using the formula for the inversion of a partitioned matrix
given in Appendix III.
Transposing the matrix gives
H
2T
AT 1 QA21 BR21 BT
5
2A21 BR21 BT
2QA21
A21
We observe that the entries of the matrix are the same as those of the
Hamiltonian matrix with some rearrangement and sign changes. Hence, we can
rearrange terms in the Hamiltonian matrix and change signs to obtain the equality
JH 5 H 2T J
where J is the matrix
J5
0
2In
In
0
(10.67)
(10.68)
By multiplying J and its transpose to obtain the identity matrix, we can verify
that the inverse of the matrix J is
J 21 5 J T 5 2J
This allows us to write the inverse of the Hamiltonian matrix as
H 21 5 ðJHJ 21 ÞT 5 JHJ T
(10.69)
429
430
CHAPTER 10 Optimal Control
Equation (10.67) allows us to examine the properties of the eigenvalues and
eigenvectors of the Hamiltonian matrix. Multiplying by any right eigenvector
gives
JHvi 5 µi Jvi 5 H 2T Jvi ; i 5 1; 2; . . .; 2n
(10.70)
where (µi, vi) are an eigenpair of the Hamiltonian matrix. Clearly, (10.70) shows
that Jv is a right eigenvector of H2T corresponding to the same eigenvalue. Since
the eigenvalues of any matrix are the same as those of its transpose, we conclude
that the Hamiltonian matrix H and its inverse H21 have the same eigenvalues.
Because the eigenvalues of a matrix are the reciprocals of those of its inverse, we
conclude that the eigenvalues of H can be written as
µi ;
1
; i 5 1; . . .; n
µi
(10.71)
If we arrange the eigenvalues of the Hamiltonian matrix so that the first n are
inside the unit circle, then the remaining n eigenvalues will be outside the unit
circle. It turns out that the stable eigenvalues of the Hamiltonian matrix are also
the closed-loop eigenvalues of the system with optimal control.
From (10.70), we also observe that Jvi,i 5 1, 2,. . ., 2n are the eigenvectors of
H2T. Premultiplying (10.70) by HT then transposing gives
µi vTi J T H 5 vTi J T ; i 5 1; 2; . . .; 2n
Using (10.69), this can be rewritten as
vTi JH 5
1 T
v J; i 5 1; 2; . . .; 2n
µi i
Thus, the left eigenvectors of H are
fvTi J; i 5 1; . . .; 2ng
Note that if vi is a right eigenvector corresponding to a stable eigenvalue, then
vTi J is a left eigenvector corresponding to an unstable eigenvalue, and vice versa.
The eigenstructure of the inverse Hamiltonian matrix governs the evolution of
the state and costate vectors. Because the closed-loop optimal control system is
stable in the steady-state provided that the stabilizability and detectability conditions are met, we expect the stable eigenvalues of the matrix to be related to the
dynamics of the closed-loop optimal control system.
THEOREM 10.1
The n stable eigenvalues of H21 are the eigenvalues of the closed-loop optimal control state
matrix Acl 5 A 2 BK with the optimal control of (10.47).
10.5 Hamiltonian system
PROOF
Let the initial state of the control system be given by the ith eigenvector vxi. Then the
response of the system will remain on the eigenvector and will be of the form (see Chapter 7)
x ðkÞ 5 µki vxi ;
and for time k 1 1
x ðk 1 1Þ 5 µk11
vxi
i
This shows that x (k), and therefore vxi is an eigenvector of Acl with eigenvalue µi for
i 5 1,. . ., 2n.
Using (10.36), we can write the costate as
λ ðkÞ 5 Sx ðkÞ 5 µki Svxi 5 µki vλi
(10.72)
where vλi 5 Svxi.
For time k 1 1, the costate is
λ ðk 1 1Þ 5 µk11
vλi
i
The two equations at time k 1 1 combine to give
vxi
x ðk 1 1Þ
5 µk11
i
vλi
λ ðk 1 1Þ
From (10.65), we also have
v
x ðk 1 1Þ
x ðkÞ
5 H 21 5 H 21 µki xi
vλi
λ ðk 1 1Þ
λ ðkÞ
Because this expression holds for any Hamiltonian system, we now have
v
v
H 21 xi 5 µi xi
vλi
vλi
In addition, multiplying by H gives
H
1 vxi
vxi
5
vλi
µi vλi
so that the eigenvectors are also those of the unstable eigenvectors of the Hamiltonian
matrix.
The eigenvector relations of the proof allow us to write the following expression for the
solution of the algebraic Riccati equation
Vλ 5 SVx
where the eigenvector matrices are given by
Vx 5 vx1 . . . vxn Vλ 5 vλ1 . . . vλn Thus, the solution of the algebraic Riccati equation is given in terms of the
stable eigenvectors of the inverse Hamiltonian matrix as
S 5 Vλ Vx21
(10.73)
431
432
CHAPTER 10 Optimal Control
We return to the optimal feedback gain of (10.47) and multiply both sides by
½R 1 BT SB
½R 1 BT SBK 5 BT SA
Rearranging terms, we can rewrite the gain in terms of the closed-loop state matrix
K 5 R21 BT SAcl
We then substitute from (10.73) for S and for Acl in terms of its eigenstructure
K 5 R21 BT Vλ Vx21 3 Vx MVx21
Acl 5 Vx MVx21 ; M 5 diagfµ1 ; . . .; µn g
We now have an expression for the optimal gain in terms of the stable eigenpairs of the
inverse Hamiltonian matrix
K 5 R21 BT Vλ MVx21
(10.74)
From Chapter 9, we know that the feedback gain that assigns a given set of
eigenvalues is unique for a single-input system but not multi-input systems. In the
latter case, there is some freedom in choosing the closed-loop eigenvectors. The
single-input case allows us to examine the effect of the weighting matrices on the
locations of the closed-loop poles.
EXAMPLE 10.9
Obtain the eigenstructure of the inverse Hamiltonian system and use it to obtain the optimal
control for the steady-state regulator for the double integrator system of Example 10.4.
The discretized double integrator system has the state equation
1 T
T 2 =2
uðkÞ
xðk 1 1Þ 5
xðkÞ 1
0 1
T
The inverse of the state matrix is
1 2T
A21 5
0
1
The inverse Hamiltonian matrix is
3
2
1
0:02 24 3 1027 23:999 3 1025
7
6
24
A 1 BR21 BT A2T Q 2BR21 BT A2T
1:004 24 3 1025
24 3 1023 7
6 4 3 10
56
H21 5
7
2T
2T
5
4
210 0
1
0
2A Q
A
0:2 21
20:02 1
Using MATLAB, we obtain the eigenvalues {1.0558 6 j0.03217, 0.9462 6 0.0310}, of which
the latter two are inside the unit circle. The corresponding eigenvectors form the matrix
3
2
5:1155 3 1023 2 j2:9492 3 1023
5:1155 3 1023 1 j2:9492 3 1023
23
23
23
23
6 29:1786 3 10 1 j16:4508 3 10
Vx
29:1786 3 10 2 j16:4508 3 10 7
7
V5
56
5
4
0:95087
0:95087
Vλ
20:014069 1 j0:3086 20:014069 2 j0:3086
Finally, we have the optimal feedback gain matrix
K 5 R21 BT Vλ MVx21 5 8:5862 5:088 Problems
Resources
Anderson, B.D.O., Moore, J.B., 1990. Optimal Control: Linear Quadratic Methods.
Prentice Hall, Englewood Cliffs, NJ.
Chong, E.D., Z_ak, S.H, 1996. An Introduction to Optimization. Wiley-Interscience,
New York.
Jacquot, R.G., 1981. Modern Digital Control Systems. Marcel Dekker, New York.
Kwakernaak, H., Sivan, R., 1972. Linear Optimal Control Systems. Wiley-Interscience,
New York.
Lewis, F.L., Syrmos, V.L., 1995. Optimal Control. Wiley-Interscience, New York.
Naidu, D.S., 2002. Optimal Control Systems. CRC Press, Boca Raton, FL.
Sage, A.P., White III, C.C., 1977. Optimum Systems Control. Prentice-Hall, Englewood
Cliffs, NJ.
PROBLEMS
10.1
Show that for a voltage source vs with source resistance Rs connected to a
resistive load RL, the maximum power transfer to the load occurs when
RL 5 Rs .
10.2
Let x be an n 3 1 vector whose entries are the quantities produced by a
manufacturer. The profit of the manufacturer is given by the quadratic
form
1
JðxÞ 5 xT Px 1 qT x 1 r
2
where P is a negative definite symmetric matrix and q and r are constant
vectors. Find the vector x to maximize profit.
(a) With no constraints on the quantity produced
(b) If the quantity produced is constrained by
Bx 5 c
where B is an m 3 n matrix, m , n, and c is a constant vector.
10.3
Prove that the rectangle of the largest area that fits inside a circle of
diameter D is a square of diagonal D.
10.4
With q 5 1 and r 5 2, S(kf) 5 1, write the design equations for a digital
optimal quadratic regulator for the integrator
x_ 5 u
The discretized statespace model
7.14 is given by
2
3 2
0:9932
20:03434
x1 ðk 1 1Þ
4 x2 ðk 1 1Þ 5 5 4 20:009456
0:9978
20:0002368 0:04994
x3 ðk 1 1Þ
10.5
of the INFANTE AUV of Problem
32
3 2
3
0
0:002988
x1 ðkÞ
0 54 x2 ðkÞ 5 1 4 20:0115 5uðkÞ
1
20:0002875
x3 ðkÞ
433
CHAPTER 10 Optimal Control
2
3
x1 ðkÞ
1 0 0 4
y1 ðkÞ
5
x2 ðkÞ 5
0 1 0
y2 ðkÞ
x3 ðkÞ
Design a steady-state linear quadratic regulator for the system using the weight
matrices Q 5 I3 and r 5 2.
10.6
A simplified linearized model of a drug delivery system to maintain
blood glucose and insulin levels at prescribed values is given by
2
3 2
32
3 2
3
20:04 24:4
0
x1 ðkÞ
1 0
x1 ðk11Þ
u ðkÞ
4 x2 ðk11Þ 5 5 4 0
20:025 1:331025 54 x2 ðkÞ 5 1 4 0 0 5 1
u2 ðkÞ
x3 ðk11Þ
x3 ðkÞ
0
0:09
0
0 0:1
2
3
x ðkÞ
y1 ðkÞ
1 0 0 4 1 5
x2 ðkÞ
5
y2 ðkÞ
0 0 1
x3 ðkÞ
where all variables are perturbations from the desired steady-state levels.1
The state variables are the blood glucose concentration x1 in mg/dl, the
blood insulin concentration x3 in mg/dl, and a variable describing the accumulation of blood insulin x2. The controls are the rate of glucose infusion u1
and the rate of insulin infusion u2, both in mg/dl/min. Discretize the system
with a sampling period T 5 5 min and design a steady-state regulator for the
system with weight matrices Q 5 I3 and R 5 2I2. Simulate the system with
the initial state xð0Þ 5 6 0 21 T , and plot the trajectory in the x1 2 x3
plane as well as the time evolution of the glucose concentration.
10.7
Optimal control problems can be solved as unconstrained optimization
problems without using Lagrange multipliers. This exercise shows the
advantages of using Lagrange multipliers and demonstrates that discretetime optimal control is equivalent to an optimization problem over the
control history.
(a) Substitute the solution of the state equation
xðkÞ 5 Ak xð0Þ 1
k21
X
Ak2i21 BuðiÞ
i50
5 Ak xð0Þ 1 CðkÞ uðkÞ
h
i
CðkÞ 5 B AB ? Ak21 B
uðkÞ 5 colfuðk 2 1Þ; . . .; uð0Þg
434
1
Chee, F., Savkin, A.V., Fernando, T.L., Nahavandi, S., 2005. Optimal HN insulin injection control
for blood glucose regulation in diabetic patients. IEEE Trans. Biomed. Eng. 52 (10), 16251631.
Problems
In the performance measure
k 21
J5
f
1 T
1X
ðxT ðkÞQðkÞxðkÞ 1 uT ðkÞRðkÞuðkÞÞ
x ðkf ÞSðkf Þxðkf Þ 1
2
2 k50
to eliminate the state vector and obtain
kf
uT ðkÞRðkÞ uðkÞ 1 2T xð0ÞðAT Þk QðkÞCðkÞ uðkÞ
1X
J5
2 k50
1 xð0ÞðAT Þk QðkÞAk xð0Þ 1 uT ðkÞRðkÞuðkÞ
!
with Qðkf Þ 5 Sðkf Þ; Rðkf Þ 5 0m3m .
(b) Without tediously evaluating the matrix Req and the vector l, explain
why it is possible to rewrite the performance measure in the equivalent form
Jeq 5
1 T
u ðkf ÞReq uðkf Þ 1 uT ðkf Þl
2
(c) Show that the solution of the optimal control problem is given by
uðkf Þ 5 2R21
eq l
10.8
For (A, B) stabilizable and (A, Q1/2) detectable, the linear quadratic regulator yields a closed-loop stable system. To guarantee that the eigenvalues
of the closed-loop system will lie inside a circle of radius 1/α, we solve
the regulator problem for the scaled state and control
xðkÞ 5 αk xðkÞ uðkÞ 5 αk11 uðkÞ
(a) Obtain the state equation for the scaled state vector.
(b) Show that if the scaled closed-loop state matrix with the optimal control uðkÞ 5 2KxðkÞ has eigenvalues inside the unit circle, then the
eigenvalues of the original state matrix with the control
uðkÞ 5 2KxðkÞ; K 5 K=α are inside a circle of radius 1/α.
10.9
Repeat Problem 10.5 with a design that guarantees that the eigenvalues
of the closed-loop system are inside a circle of radius equal to one-half.
10.10
Show that the linear quadratic regulator with cross-product term of the form
J 5 xT ðkf ÞSðkf Þxðkf Þ 1
1uT ðkÞRðkÞuðkÞ
kX
f 21
k5k0
ðxT ðkÞQðkÞxðkÞ 1 2xT Su
435
436
CHAPTER 10 Optimal Control
is equivalent to a linear quadratic regulator with no cross-product term
with the cost
J 5 xT ðkf ÞSðkf Þxðkf Þ 1
kX
f 21
ðxT ðkÞQðkÞxðkÞ
k5k0
1uT ðkÞRðkÞuT ðkÞ
Q 5 Q 2 SR21 ST
uðkÞ 5 uðkÞ 1 R21 ST xðkÞ
and the plant dynamics
xðk 1 1Þ 5 AxðkÞ 1 BuðkÞ
A 5 A 2 BR21 BT
10.11
k 5 k0 ; . . .; kf 2 1
Rewrite the performance measure shown in Problem 10.10 in terms of a
combined input and state vector col{x(k), u(k)}. Then use the
Hamiltonian to show that for the linear quadratic regulator with crossproduct term, a sufficient condition for a minimum is that the matrix
Q N
NT R
must be positive definite.
COMPUTER EXERCISES
10.12
Design a steady-state regulator for the INFANTE AUV presented in
Problem 10.5 with the performance measure modified to include a crossproduct term with
S5 1
0:2
0:1 T
(a) Using the MATLAB command dlqr and the equivalent problem with
no cross-product as in Problem 10.10
(b) Using the MATLAB command dlqr with the cross-product term
10.13
Design an output quadratic regulator for the INFANTE UAV presented in
Problem 10.5 with the weights Qy 5 1 and r 5 100. Plot the output
response for the initial condition vector x(0) 5 [1, 0, 1]T.
10.14
Design an optimal LQ statespace tracking controller for the drug delivery system presented in Problem 10.6 to obtain zero steady-state error
due to a unit step input.
Computer Exercises
10.15
Write a MATLAB script that determines the steady-state quadratic regulator for the inertial system of Example 10.4 for r 5 1 and different values
of Q. Use the following three matrices, and discuss the effect of changing
Q on the results of your computer simulation:
1 0
(a) Q 5
0 1
10 0
(b) Q 5
0 10
100
0
(c) Q 5
0
100
10.16
The linearized analog state-space model of the three tanks system shown
in Figure P10.16 can be written as2
2
3 2
32
3 2 3
21
0
0
x1 ðtÞ
1
x_1 ðtÞ
4 x_2 ðtÞ 5 5 4 1
5
4
5
4
21
0
x2 ðtÞ 1 0 5uðtÞ
0
1
21
0
x_3 ðtÞ
x3 ðtÞ
2
2
3
x1 ðtÞ
yðtÞ 5 4 0 0 1 4 x2 ðtÞ 5
x3 ðtÞ
where xi(t) is the level of the ith tank and u(t) is the inlet flow of the first
tank. A digital controller is required to control the fluid levels in the
tanks. The controller must provide a fast step response without excessive
overshoot that could lead to fluid overflow. Using the appropriate
u
x1
x2
x3
FIGURE P10.16
Schematic of the three tanks system.
2
Koenig, D.M., 2009. Practical Control Engineering. McGraw-Hill, New York.
437
438
CHAPTER 10 Optimal Control
MATLAB commands, design a digital LQ tracking controller for the system, and then simulate the closed-loop dynamics using SIMULINK.
Simulate the closed-loop system using SIMULINK. Use a sampling
period of 0.1 s, a state-weighting matrix Q 5 0.1 I4, and an input-weighting matrix R 5 1.
10.17
For the system in Problem 10.15, write a MATLAB program that plots
the locus of the eigenvalues of the closed-loop system for R 5 1 and
Q 5 q I4 with q in the range 0.1 # q # 10.
CHAPTER
Elements of Nonlinear
Digital Control Systems
11
OBJECTIVES
After completing this chapter, the reader will be able to do the following:
1. Discretize special types of nonlinear systems.
2. Determine the equilibrium point of a nonlinear discrete-time system.
3. Classify equilibrium points of discrete-time systems based on their state plane
trajectories.
4. Determine the stability or instability of nonlinear discrete-time systems.
5. Design controllers for nonlinear discrete-time systems.
Most physical systems do not obey the principle of superposition and can therefore
be classified as nonlinear. By limiting the operating range of physical systems, it is
possible to approximately model their behavior as linear. This chapter examines
the behavior of nonlinear discrete systems without this limitation. We begin by
examining the behavior of nonlinear continuous-time systems with piecewise constant inputs. We discuss Lyapunov stability theory and input-output stability both
for nonlinear and linear systems. We provide a simple controller design based on
Lyapunov stability theory.
11.1 Discretization of nonlinear systems
Discrete-time models are easily obtained for linear continuous-time systems from
their transfer functions or the solution of the state equations. For nonlinear systems, transfer functions are not defined, and the state equations are analytically
solvable in only a few special cases. Thus, it is no easy task to obtain discretetime models for nonlinear systems.
We examine the nonlinear differential equation
x_ 5 fðxÞ 1 BðxÞu
(11.1)
where x and f are n 3 1 vectors, u is an m 3 1 vector, B(x) is a n 3 m matrix, and
B(.) and f(.) are continuous functions of all their arguments. We assume that the
input is defined by
uðtÞ 5 uðkTÞ 5 uðkÞ; tA½kT; ðk 1 1ÞT
(11.2)
439
440
CHAPTER 11 Elements of Nonlinear Digital Control Systems
For each sampling period we have the model
x_ 5 fðxÞ 1 BðxÞuðkÞ
(11.3)
where k 5 0, 1, 2, . . . . The behavior of the discrete-time system can, in theory,
be obtained from the solution of equations of the form (11.3) with the appropriate
initial conditions. In practice, only a numerical solution is possible except in
a few special cases. One way to obtain a solution for nonlinear systems is to
transform the dynamics of the system to obtain equivalent linear dynamics. The
general theory governing such transformations, known as global or extended
linearization, is beyond the scope of this text. However, in some special cases it
is possible to obtain equivalent linear models and use them for discretization quite
easily without resorting to the general theory. These include models that occur
frequently in some applications. We present four cases where extended linearization is quite simple.
11.1.1 Extended linearization by input redefinition
Consider the nonlinear model
_ 5 vðtÞ
MðqÞ€q 1 mðq; qÞ
(11.4)
where M(q) is an invertible m 3 m matrix, and m, q, and v are m 3 1 vectors.
A natural choice of state variables for the system is the 2m 3 1 vector
_
x 5 colfx1 ; x2 g 5 colfq; qg
(11.5)
To obtain equivalent linear dynamics for the system, we redefine the input
vector as
uðtÞ 5 q
€
(11.6)
This yields the set of double integrators with state equations
x_ 1 5 x2
x_ 2 5 uðtÞ
_
uðtÞ 5 M 21 ðqÞ½vðtÞ 2 mðq; qÞ
(11.7)
The solution of the state equations can be obtained by Laplace transformation,
and it yields the discrete-state equation
xðk 1 1Þ 5 Ad xðkÞ 1 Bd uðkÞ
with
Ad 5
Im
0
TIm
Im
Bd 5
T 2 =2 Im
TIm
(11.8)
(11.9)
11.1 Discretization of nonlinear systems
With digital control, the input is piecewise constant and is obtained using the
system model
vðkÞ 5 Mðx1 ðkÞÞuðkÞ 1 mðx1 ðkÞ; x2 ðkÞÞ
(11.10)
The expression for the input v is approximate because it assumes that the state
does not change appreciably over a sampling period with a fixed acceleration
input. The approximation is only acceptable for sufficiently slow dynamics.
The model shown in (11.4) includes many important physical systems. In
particular, a large class of mechanical systems, including the m-D.O.F. (degreeof-freedom) manipulator presented in Example 7.4, are in the form (11.4). Recall
that the manipulator is governed by the equation of motion
_ q_ 1 gðqÞ 5 τ
MðqÞ€q 1 Vðq; qÞ
(11.11)
where
q 5 vector of generalized coordinates
M(q) 5 m 3 m positive definite inertia matrix
_ 5 m 3 m matrix of velocity-related terms
Vðq; qÞ
g(q) 5 m 3 1 vector of gravitational terms
τ 5 m 3 1 vector of generalized forces
Clearly, the dynamics of the manipulator are a special case of (11.4). As
applied to robotic manipulators, extended linearization by input redefinition is
known as the computed torque method. This is because the torque is computed from the acceleration if measurements of positions and velocities are
available.
EXAMPLE 11.1
Find an approximate discrete-time model for the 2-D.O.F. robotic manipulator described in
Example 7.5 and find an expression for calculating the torque vector.
Solution
The manipulator dynamics are governed by
2
ðm1 1 m2 Þl21 1 m2 l22 1 2m2 l1 l2 cosðθ2 Þ
MðθÞ 5 4
m2 l22 1 m2 l1 l2 cosðθ2 Þ
m2 l22 1 m2 l1 l2 cosðθ2 Þ
m2 l22
2
3
2m2 l1 l2 sinðθ2 Þθ_2 ð2θ_1 1 θ_2 Þ
_ θ_ 5 4
5
Vðθ; θÞ
2
m2 l1 l2 sinðθ2 Þθ_
1
2
gðθÞ 5 4
ðm1 1 m2 Þgl1 sinðθ1 Þ 1 m2 gl2 sinðθ1 1 θ2 Þ
m2 gl2 sinðθ1 1 θ2 Þ
3
5
3
5
441
442
CHAPTER 11 Elements of Nonlinear Digital Control Systems
For this system, the coordinate vector is q 5 [θ1 θ2]T. We have a fourth-order linear system
of the form
x_ 1 ðtÞ 5 x2 ðtÞ
x_ 2 ðtÞ 5 uðtÞ
x1 ðtÞ 5 q 5 ½ x11 ðtÞ x12 ðtÞ T
x2 ðtÞ 5 q_ 5 ½ x21 ðtÞ x22 ðtÞ T
uðtÞ 5 ½ τ 1 ðtÞ τ 2 ðtÞ T
As in (11.10), the torque is calculated using the equation
τðkÞ 5 Mðx1 ðkÞÞuðkÞ 1 Vðx1 ðkÞ; x2 ðkÞÞx2 ðkÞ 1 gðx1 ðkÞÞ
11.1.2 Extended linearization by input and state redefinition
Consider the nonlinear state equations
z_ 1 5 f 1 ðz1 ; z2 Þ
z_ 2 5 f 2 ðz1 ; z2 Þ 1 Gðz1 ; z2 ÞvðtÞ
(11.12)
where fi(.), i 5 1, 2, v are m 3 1, and G is m 3 m. We redefine the state variables and
input as
x1 5 z 1
x2 5 z_ 1
uðtÞ 5 z€ 1
(11.13)
The new variables have the linear dynamics of (11.7) and the discrete-time model
of (11.811.9).
Using the state equations, we can rewrite the new input as
uðtÞ 5
@f 1 ðz1 ; z2 Þ
@f 1 ðz1 ; z2 Þ
½f 2 ðz1 ; z2 Þ 1 Gðz1 ; z2 ÞvðtÞ
f 1 ðz1 ; z2 Þ 1
@z1
@z2
We solve for the nonlinear system input at time k to obtain
2
321
@f
ðz
;
z
Þ
1 1
2
vðkÞ 5 4
Gðz1 ; z2 Þ5
@z2
9
8
=
<
@f 1 ðz1 ; z2 Þ
@f 1 ðz1 ; z2 Þ
3 uðkÞ 2
f 1 ðz1 ; z2 Þ 2
f 2 ðz1 ; z2 Þ
;
:
@z1
@z2
(11.14)
(11.15)
The expression for the input v(k) is approximate because the state of the
system changes over a sampling period for a fixed input u(k).
11.1 Discretization of nonlinear systems
EXAMPLE 11.2
Discretize the nonlinear differential equation
1
z_1 5 2 ðz1 z2 Þ2
2
z_2 5 4z1 z32 1 1:5z2 2 v;
z1 6¼ 0; z2 6¼ 0
Solution
We differentiate the first state equation to obtain
z€1 5 2z1 z22 z_1 2 z21 z2 z_2
5 23:5z31 z42 2 1:5z21 z22 1 z21 z2 v 5 uðtÞ;
z1 6¼ 0; z2 6¼ 0
We have the linear model
x_1 5 x2
x_2 5 uðtÞ
This gives the equivalent discrete-time model
1 T
xðkÞ 1
xðk 1 1Þ 5
0 1
T 2 =2
T
uðkÞ
with x(k) 5 [x1(k) x2(k)]T 5 [z1(k) z1(k 1 1)]T. If measurements of the actual state of the
system are available, then the input is computed using
vðkÞ 5 3:5z1 ðkÞz32 ðkÞ 1 1:5z2 ðkÞ 1
uðkÞ
;
z21 ðkÞz2 ðkÞ
z1 6¼ 0; z2 6¼ 0
11.1.3 Extended linearization by output differentiation
Consider the nonlinear statespace model
z_ 5 fðzÞ 1 GðzÞvðtÞ
y 5 cðzÞ
(11.16)
where z and f are n 3 1, v is m 3 1, y and c are l 3 1, and G is m 3 m. If we
differentiate each scalar output equation and substitute from the state equation,
we have
dyi
@cT ðzÞ
5 i
z_
dt
@z
@cT ðzÞ
½fðzÞ 1 GðzÞvðtÞ
i 5 1; . . . ; l
(11.17)
5 i
@z
We denote the operation of partial differentiation with regard to a vector and
the multiplication by its derivative by
@y
Lx y 5
x_
(11.18)
@x
443
444
CHAPTER 11 Elements of Nonlinear Digital Control Systems
Hence, the time derivative of the ith output is equivalent to the operation Lz_ yi .
We denote ri repetitions of this operation by the symbol Lri i yi .
If the coefficient of the input vector v
@cTi ðzÞ
GðzÞ 5 0T
(11.19)
@z
is nonzero, we stop to avoid input differentiation. If the coefficient is zero, we
repeat the process until the coefficient becomes nonzero. We define an input equal
to the term where the input appears and write the state equations in the form
d di yi
5 ui ðtÞ
dtdi
(11.20)
ui ðtÞ 5 Lri i yi ;
i 5 1; . . .; l
where the number of derivatives ri required for the input to appear in the expression is known as the ith relative degree. The linear system is in the form of l sets
of integrators, and its order is
nl 5
n
X
ri # n
(11.21)
i51
Because the order of the linear model can be less than the original order of the
nonlinear system n, the equivalent linear model of the nonlinear system can have
unobservable dynamics. The unobservable dynamics are known as the internal
dynamics or zero dynamics. For the linear model to yield an acceptable discretetime representation, we require the internal dynamics to be stable and sufficiently
fast so as not to significantly alter the time response of the system. Under these
conditions, the linear model provides an adequate though incomplete description of
the system dynamics, and we can discretize each of the l linear subsystems to
obtain the overall discrete-time model.
EXAMPLE 11.3
Consider a model of drug receptor binding in a drug delivery system. The drug is assumed
to be divided between four compartments in the body. The concentrations in the four
compartments are given by
z_1 5 a11 z1 1 a12 z1 z2 1 b1 d
z_2 5 a21 z2 1 a22 z1 z2 1 b2 d
z_3 5 a31 z1 1 a32 z3
z_4 5 a41 z2
where d is the drug dose, zi is the concentration in the ith compartment, and aij are timevarying coefficients. The second compartment represents the drug receptor complex.
Hence, the output equation is given by
y 5 z2
Obtain an equivalent linear model by output differentiation then use it to obtain an
approximate discrete-time model of the system.
11.1 Discretization of nonlinear systems
Solution
We differentiate the output equation once:
y_ 5 z_2 5 a21 z2 1 a22 z1 z2 1 b2 d
The derivative includes the input, and no further differentiation is possible because it
would lead to input derivative terms.
The model of the linearized system is
x_ 5 uðtÞ
uðtÞ 5 a21 z2 1 a22 z1 z2 1 b2 d
The model is first order, even though we have a third-order plant, and it is only equivalent
to part of the dynamics of the original system. We assume the system is detectable with
fast unobservable dynamics, which is reasonable for the drug delivery system. Hence, the
linear model provides an adequate description of the system, and we obtain the discretetime model
xðk 1 1Þ 5 xðkÞ 1 TuðkÞ
The drug dose is given by
dðkÞ 5
1
½uðkÞ 2 a21 z2 ðkÞ 1 a22 z1 ðkÞz2 ðkÞ
b2
11.1.4 Extended linearization using matching conditions
Theorem 11.1 gives the analytical solution in a special case where an
equivalent linear model can be obtained for a nonlinear system by a simple state
transformation.
THEOREM 11.1
For the system of (11.1), let B(x) and f(x) satisfy the matching conditions
BðxÞ 5 B1 ðxÞB2
fðxÞ 5 B1 ðxÞAhðxÞ
(11.22)
where B1(x) is an n 3 n matrix invertible in some region D, B2 is an n 3 m constant vector, A is
an n 3 n matrix, and h (x) has the derivative given by
@hðxÞ
5 ½B1 ðxÞ21
@x
(11.23)
with h(.) invertible in the region D. Then the solution of (11.1) over D is
xðtÞ 5 h21 ðwðtÞÞ
(11.24)
where w is the solution of the linear equation
_ 5 AwðtÞ 1 B2 uðtÞ
wðtÞ
(11.25)
445
446
CHAPTER 11 Elements of Nonlinear Digital Control Systems
PROOF
From (11.24), w 5 h(x). Differentiating with respect to time and substituting from (11.1)
and (11.22) gives
_5
w
@hðxÞ
x_ 5 ½B1 ðxÞ21 ffðxÞ 1 B1 ðxÞB2 ug
@x
5 AhðxÞ 1 B2 u
5 Aw 1 B2 u
Note that, if the decomposition of Theorem 11.1 is possible, we do not need
to solve a partial differential equation to obtain the vector function h(x), but we
do need to find its inverse function to obtain the vector x. The solution of a partial differential equation is a common requirement for obtaining transformations
of nonlinear systems to linear dynamics that can be avoided in special cases.
The solution of the state equation for w over any sampling period T is
wðk 1 1Þ 5 Aw wðkÞ 1 Bw uðkÞ
where
Aw 5 e ;
Bw 5
AT
ðT
eAτ B2 dτ
0
This is a discrete statespace equation for the original nonlinear system.
EXAMPLE 11.4
Discretize the nonlinear differential equation
1
x_1
5
x_2
2
2
2 x1 =x2
1
3
8x2 =x1 1 3x2
0
3 u;
2x2
x1 6¼ 0; x2 6¼ 0
Solution
Assuming a piecewise constant input u, we rewrite the nonlinear equation as
x_1
x2
5 1
x_2
0
0
2x32
0
24
21
3
0
2x21
1
1
uðkÞ
1
x22
2 =2
To apply Theorem 11.1, we define the terms
"
B1 ðxÞ 5
"
A5
#
x21
0
0
2x32
0
21
24
3
B2 5
" #
0
"
#
hðxÞ 5
1
2x21
1
x22
2 =2
#
11.2 Nonlinear difference equations
We verify that the Jacobian of the vector satisfies
@hðxÞ
@h1 =@x1 @h1 =@x2
x22
5 ½B1 ðxÞ21 5 1
5
@h2 =@x1 @h2 =@x2
0
@x
0
2x23
2
and that Theorem 11.1 applies.
We first solve the linear state equation
w_ 1
0
21 w1
0
5
1
uðkÞ
w_ 2
w2
24
3
1
to obtain
1
w1 ðk 1 1Þ
1
21 4T
e 1
5
24
4
w2 ðk 1 1Þ
5
4T
1
21 e 2 1
1
1
4
5
4
4 1 2T
w1 ðkÞ
e
4 1
w2 ðkÞ
1
ð1 2 e2T Þ uðkÞ
1
where w(k) 5 h[x(k)], k 5 0, 1, 2, . . . We now have the recursion
1
4 1 2T
1
21 4T
w1 ðkÞ
w1 ðk 1 1Þ
e
e 1
5
4 1
24
4
w2 ðk 1 1Þ
w2 ðkÞ
5
1
1
5
1
21 e4T 2 1
ð1 2 e2T Þ uðkÞ
1
1
4
4
Using the inverse of the function h[x(k)], we solve for the state
x1 ðk 1 1Þ 5 2 1=w1 ðk 1 1Þ;
x2 ðk 1 1Þ 5 6 1=ð2w2 ðk11ÞÞ1=2
which is clearly nonunique. However, one would be able to select the appropriate solution
because changes in the state variable should not be large over a short sampling period. For
a given sampling period T, we can obtain a discrete nonlinear recursion for x(k).
It is obvious from Example 11.4 that the analysis of discrete-time systems based
on nonlinear analog systems is only possible in special cases. In addition, the transformation is sensitive to errors in the system model. We conclude that this is often not
a practical approach to controlling system design. However, most nonlinear control
systems are today digitally implemented even if based on an analog design. We therefore return to a discussion of system (11.1) with the control (11.2) in Section 11.5.
11.2 Nonlinear difference equations
For a nonlinear system with a DAC and ADC, system identification can yield a
discrete-time model directly. A discrete-time model can also be obtained analytically in a few special cases, as discussed in Section 11.1. We thus have a nonlinear
difference equation to describe a nonlinear discrete-time system. Unfortunately,
nonlinear difference equations can be solved analytically in only a few special
cases. We discuss one special case where such solutions are available.
447
448
CHAPTER 11 Elements of Nonlinear Digital Control Systems
11.2.1 Logarithmic transformation
Consider the nonlinear difference equation
½yðk1nÞαn ½yðk1n21Þαn21 . . . ½yðkÞα0 5 uðkÞ
(11.26)
where αi, i 5 0, 1, . . ., n are constant. Then, taking the natural log of (11.26) gives
αn xðk 1 nÞ 1 αn21 xðk 1 n 2 1Þ . . . 1 α0 xðkÞ 5 vðkÞ
(11.27)
with x(k 1 i) 5 ln[y(k 1 i)], i 5 0, 1, 2, . . . , n, and v(k) 5 ln[u(k)]. This is a linear
difference equation that can be easily solved by z-transformation. Finally, we obtain
yðkÞ 5 exðkÞ
(11.28)
EXAMPLE 11.5
Solve the nonlinear difference equation
½yðk 1 2Þ½yðk11Þ5 ½yðkÞ4 5 uðkÞ
with zero initial conditions and the input
uðkÞ 5 e210k
Solution
Taking the natural log of the equation, we obtain
xðk 1 2Þ 1 5xðk 1 1Þ 1 4xðkÞ 5 210k
The z-transform of the equation with zero initial conditions
z2 1 5z 1 4 XðzÞ 5
210z
ðz21Þ2
yields X(z) as
XðzÞ 5
210z
ðz21Þ2 ðz 1 1Þðz 1 4Þ
Inverse z-transforming gives the discrete-time function
xðkÞ 5 2k 1 0:7 2 0:8333ð21Þk 1 0:1333ð24Þk ;
k$0
Hence, the solution of the nonlinear difference equation is
yðkÞ 5 e2k10:720:8333ð21Þ 10:1333ð24Þ ;
k
k
k$0
11.3 Equilibrium of nonlinear discrete-time systems
Equilibrium is defined as a condition in which a system remains indefinitely
unless it is disturbed. If a discrete-time system is described by the nonlinear difference equation
xðk 1 1Þ 5 f½xðkÞ 1 B½xðkÞuðkÞ
(11.29)
11.3 Equilibrium of nonlinear discrete-time systems
then at an equilibrium, it is governed by the identity
xe 5 f½xe 1 B½xe uðkÞ
(11.30)
where xe denotes the equilibrium state. We are typically interested in the
equilibrium for an unforced system, and we therefore rewrite the equilibrium
condition (11.30) as
xe 5 f½xe (11.31)
In mathematics, such an equilibrium point is known as a fixed point.
Note that the behavior of the system in the vicinity of its equilibrium point
determines whether we classify the equilibrium as stable or unstable. For a stable
equilibrium, we expect the trajectories of the system to remain arbitrarily close to or
to converge to the equilibrium. Unlike continuous-time systems, convergence to an
equilibrium point can occur after a finite time period.
Clearly, both (11.30) and (11.31), in general, have more than one solution so
that nonlinear systems often have several equilibrium points. This is demonstrated
in Example 11.6.
EXAMPLE 11.6
Find the equilibrium points of the nonlinear discrete-time system
x ðkÞ
x1 ðk 1 1Þ
5 23
x1 ðkÞ
x2 ðk 1 1Þ
Solution
The equilibrium points are determined from the condition
x ðkÞ
x1 ðkÞ
5 23
x1 ðkÞ
x2 ðkÞ
or equivalently,
x2 ðkÞ 5 x1 ðkÞ 5 x31 ðkÞ
Thus, the system has the three equilibrium points: (0, 0), (1, 1), and (21, 21).
To characterize equilibrium points, we need the following definition.
DEFINITION 11.1
Contraction. A function f(x) is known as a contraction if it satisfies
:fðx 2 yÞ: # α:x 2 y:;
jαj , 1
(11.32)
where α is known as a contraction constant and jj.jj is any vector norm.
Theorem 11.2 provides conditions for the existence of an equilibrium point for a
discrete-time system. The proof is left as an exercise.
449
450
CHAPTER 11 Elements of Nonlinear Digital Control Systems
THEOREM 11.2
A contraction f(x) has a unique fixed point.
EXAMPLE 11.7
Determine if the following nonlinear discrete-time system converges to the origin:
xðk 1 1Þ 5 af ½xðkÞ; k 5 0; 1; 2; . . .
f ½xðkÞ , xðkÞ ; jaj , 1
jaj
Solution
We have the inequality
xðk 1 1Þ 5 af ½xðkÞ , jaj xðkÞ # αxðkÞ;
k 5 0; 1; 2; . . . ; α , 1
jaj
Hence, the system converges to a fixed point at the origin as k-N.
If a continuous-time system is discretized, then all its equilibrium points will
be equilibrium points of the discrete-time system. This is because discretization
corresponds to the solution of the differential equation governing the original system at the sampling points. Because the analog system at an equilibrium remains
there, the discretized system will have the same behavior and therefore the same
equilibrium.
11.4 Lyapunov stability theory
Lyapunov stability theory is based on the idea that at a stable equilibrium, the
energy of the system has a local minimum, whereas at an unstable equilibrium, it
is at a maximum. This property is not restricted to energy and is in fact shared by
a class of function that depends on the dynamics of the system. We call such
functions Lyapunov functions.
11.4.1 Lyapunov functions
If a Lyapunov function can be found for an equilibrium point, then it can be used
to determine its stability or instability. This is particularly simple for linear systems but can be complicated for a nonlinear system.
We begin by examining the properties of energy functions that we need to
generalize and retain for a Lyapunov function. We note the following:
•
•
Energy is a nonnegative quantity.
Energy changes continuously with its arguments.
We call functions that are positive except at the origin positive definite. We provide
a formal definition of this property.
11.4 Lyapunov stability theory
DEFINITION 11.2
A scalar continuous function V(x) is said to be positive definite if
•
•
V(0) 5 0.
V(x) . 0 for any nonzero x.
Similar definitions are also useful where the greater-than sign in the second condition is replaced by other inequalities with all other properties unchanged. Thus, we
can define positive semidefinite ($), negative definite (,), and negative semidefinite (#) functions. If none of these definitions apply, the function is said to be indefinite. Definition 11.2 may hold locally, and then the function is called locally positive
definite, or globally, in which case it is globally positive definite. Similarly, we characterize other properties, such as negative definiteness, as local or global.
A common choice of definite function is the quadratic form xTPx. The sign of
the quadratic form is determined by the eigenvalues of the matrix P (see Appendix
III). The quadratic form is positive definite if the matrix P is positive definite, in
which case its eigenvalues are all positive. Similarly, we can characterize the quadratic form as negative definite if the eigenvalues of P are all negative, positive
semidefinite if the eigenvalues of P are positive or zero, negative semidefinite if
negative or zero, and indefinite if P has positive and negative eigenvalues.
Quadratic forms are a common choice of Lyapunov function because of their simple mathematical properties. However, it is often preferable to use other Lyapunov
functions with properties tailored to suit the particular problem. We now list the
mathematical properties of Lyapunov functions.
DEFINITION 11.3
A scalar function V(x) is a Lyapunov function in a region D if it satisfies the following conditions in D:
•
•
It is positive definite.
It decreases along the trajectories of the system—that is,
∆VðkÞ 5 Vðxðk 1 1ÞÞ 2 VðxðkÞÞ , 0;
k 5 0; 1; 2; . . .
(11.33)
Definition 11.3 is used in local stability theorems. To prove global stability,
we need an additional condition in addition to extending the two listed earlier.
DEFINITION 11.4
A scalar function V(x) is a Lyapunov function if it satisfies the following conditions:
•
•
•
It is positive definite.
It decreases along the trajectories of the system.
It is radially unbounded (i.e., it uniformly satisfies).
VðxðkÞÞ-N;
as :xðkÞ:-N
(11.34)
451
452
CHAPTER 11 Elements of Nonlinear Digital Control Systems
This last condition ensures that whenever the function V remains bounded, the
state vector will also remain bounded. The condition must be satisfied regardless
of how the norm of the state vector grows unbounded so that a finite value of V
can always be associated with a finite state. Note that unlike Definition 11.3,
Definition 11.4 requires that the first and second conditions be satisfied globally.
11.4.2 Stability theorems
We provide some sufficient conditions for the stability of nonlinear discrete-time
systems, then we specialize our results to linear time-invariant systems.
THEOREM 11.3
The equilibrium point x 5 0 of the nonlinear discrete-time system
xðk 1 1Þ 5 f½xðkÞ;
k 5 0; 1; 2; . . .
(11.35)
is asymptotically stable if there exists a locally positive definite Lyapunov function for the
system satisfying Definition 11.3.
PROOF
For any motion along the trajectories of the system, we have
∆VðkÞ 5 Vðxðk 1 1ÞÞ 2 VðxðkÞÞ
5 Vðf½xðkÞÞ 2 Vðf½xðk 2 1ÞÞ , 0;
k 5 0; 1; 2; . . .
This implies that as the motion progresses, the value of V decreases continuously. However,
because V is bounded below by zero, it converges to zero. Because V is only zero for zero
argument, the state of the system must converge to zero.
EXAMPLE 11.8
Investigate the stability of the system using the Lyapunov stability approach:
x1 ðk 1 1Þ 5 0:2x1 ðkÞ 2 0:08x22 ðkÞ
x2 ðk 1 1Þ 5 20:3x1 ðkÞx2 ðkÞ;
k 5 0; 1; 2; . . .
Solution
We first observe that the system has an equilibrium point at the origin. We select the
quadratic Lyapunov function
VðxÞ 5 xT Px; P 5 diagfp1 ; 1g
Note that the unity entry simplifies the notation without affecting the results because the
form of the function is unchanged if it is multiplied by any positive scalar p2.
The corresponding difference is
∆VðkÞ 5 p1 f½0:2x1 ðkÞ20:08x22 ðkÞ2 2 x21 ðkÞg 1 f½20:3x1 ðkÞx2 ðkÞ2 2 x22 ðkÞg
5 20:96p1 x21 ðkÞ 1 f0:09x21 ðkÞ 2 0:032p1 x1 ðkÞ 1 0:0064p1 x22 ðkÞ 2 1gx22 ðkÞ
k 5 0; 1; 2; . . .
11.4 Lyapunov stability theory
The difference remains negative provided that the term between the braces is negative.
We restrict the magnitude of x2(k) to less than 12.5 (the square root of the reciprocal
of 0.0064) and then simplify the condition to 9x21(k) 2 3.2p1x1(k) 1 100(p1 2 1) , 0,
k 5 0,1,2, . . .
Choosing a very small but positive value for p1 makes the two middle terms negligible. This leaves two terms that are negative for values of x1 of magnitude smaller than
3.33. For example, a plot of the LHS of the inequality for p1 verifies that it is negative
in the selected x1 range. Thus, the difference remains negative for initial conditions
that are inside a circle of radius approximately equal to 3.33 centered at the origin. By
Theorem 11.3, we conclude that the system is asymptotically stable in the region
jjxjj , 3.33.
The system is actually stable outside this region, but our choice of Lyapunov function
cannot provide a better estimate of the stability region than the one we obtained.
For global asymptotic stability, the system must have a unique equilibrium
point. Otherwise, starting at an equilibrium point prevents the system from converging to another equilibrium state. We can therefore say that a system is globally
asymptotically stable and not just its equilibrium point. The following theorem
gives a sufficient condition for global asymptotic stability.
THEOREM 11.4
The nonlinear discrete-time system
xðk 1 1Þ 5 f½xðkÞ;
k 5 0; 1; 2; . . .
(11.36)
with equilibrium x(0) 5 0 is globally asymptotically stable if there exists a globally positive
definite, radially unbounded Lyapunov function for the system satisfying Definition 11.4.
PROOF
The proof is similar to that of Theorem 11.3 and the details are omitted. The radial unboundedness condition guarantees that the state of the system will converge to zero with the
Lyapunov function.
The results of this section require the difference ∆V of the Lyapunov function
to be negative definite. In some cases, this condition can be relaxed to negative
semidefinite. In particular, if the nonzero values of the vector x for which ∆V is
zero are ones that are never reached by the system, then they have no impact on
the stability analysis. This leads to the following result.
COROLLARY 11.1
The equilibrium point x 5 0 of the nonlinear discrete-time system of (11.36) is asymptotically stable if there exists a locally positive definite Lyapunov function with negative semidefinite difference ∆V(k) for all k for the system and with ∆V(k) zero only for x 5 0.
453
454
CHAPTER 11 Elements of Nonlinear Digital Control Systems
Note that the preceding theorems only provide a sufficient stability condition.
Thus, failure of the stability test does not prove instability. In the linear timeinvariant case, a much stronger result is available.
11.4.3 Rate of convergence
In some cases, we can use a Lyapunov function to determine the rate of convergence of the system to the origin. In particular, if we can rewrite the difference of
the Lyapunov function in the form
∆VðkÞ # 2αVðxðkÞÞ;
0,α,1
(11.37)
substituting for the difference gives the recursion
Vðxðk 1 1ÞÞ # ð1 2 αÞVðxðkÞÞ
(11.38)
The upper bound of the constant α guarantees that the coefficient of V is positive. The solution of the difference equation is
VðxðkÞÞ # ð12αÞk Vðxð0ÞÞ
(11.39)
If V is a Lyapunov function, then it converges to zero with the state of the system and its rate of convergence allows us to estimate the rate of convergence to
the equilibrium point. If in addition V is a quadratic form, then the rate of convergence of the state is the square root of that of V.
EXAMPLE 11.9
Suppose that the quadratic form xTx is a Lyapunov function for a discrete-time system
with difference
2
4
∆VðkÞ 5 20:25:xðkÞ: 2 0:5:xðkÞ: , 0
Characterize the convergence of the system trajectories to the origin.
Solution
2
4
2
∆VðkÞ 5 2 0:25:xðkÞ: 2 0:5:xðkÞ: # 20:25:xðkÞ: 5 0:25VðxðkÞÞ
VðxðkÞÞ , ð120:25Þk Vðxð0ÞÞ
2
:xðkÞ: , 0:75k :xð0Þ:
2
:xðkÞ: # 0:866 :xð0Þ:
k
The trajectories of the system converge to the origin exponentially with convergence rate
faster than 0.86.
11.4.4 Lyapunov stability of linear systems
Lyapunov stability results typically provide us with sufficient conditions. Failure
to meet the conditions of a Lyapunov test leaves us with no conclusion and with
11.4 Lyapunov stability theory
the need to repeat the test using a different Lyapunov function or to try a different
test. For linear systems, Lyapunov stability can provide us with necessary and
sufficient stability conditions.
THEOREM 11.5
The linear time-invariant discrete-time system
k 5 0; 1; 2; . . .
xðk 1 1Þ 5 Ad xðkÞ;
(11.40)
is asymptotically stable if and only if for any positive definite matrix Q, there exists a unique
positive definite solution P to the discrete Lyapunov equation
ATd PAd 2 P 5 2Q
(11.41)
PROOF
We drop the subscript d for brevity.
Sufficiency
Consider the Lyapunov function
VðxðkÞÞ 5 xT ðkÞPxðkÞ
with P a positive definite matrix. The change in the Lyapunov function along the trajectories
of the system is
∆VðkÞ 5 Vðxðk 1 1ÞÞ 2 VðxðkÞÞ
5 xT ðkÞ½AT PA 2 PxðkÞ 5 2 xT ðkÞQxðkÞ , 0
Hence, the system is stable by Theorem 11.3.
Necessity
We first show that the solution of the Lyapunov equation is given by
P5
N
X
ðAT Þk QAk
k50
This is easily verified by substitution in the Lyapunov equation and then changing the index
of summation as follows:
AT
"
N
X
#
ðAT Þk QAk A 2
k50
N
X
k50
ðAT Þk QAk 5
N
X
ðAT Þj QAj 2
j51
N
X
ðAT Þk QAk 5 2Q
k50
To show that for any positive definite Q, P is positive definite, consider the quadratic form
xT Px 5
N
N
X
X
xT ðAT Þk QAk x 5 xT Qx 1
yTk Qyk
k50
k51
For positive definite Q, the first term is positive for any nonzero x and the other terms are
nonnegative. It follows that P is positive definite.
455
456
CHAPTER 11 Elements of Nonlinear Digital Control Systems
Let the system be stable but for some positive definite Q there is no finite solution P to
the Lyapunov equation. We show that this leads to a contradiction.
Recall that the state-transition matrix of the discrete system can be written as
Ak 5
n
X
Zi λki
(11.42)
k51
Let λmax be spectral radius of the state matrix and let the largest norm of its constituent
matrices be :Z:max. We use any matrix norm to obtain the inequality
:Ak : 5 :ðAT Þk : # n:Z:max λkmax 5 αλkmax ;
k$0
For a stable system, we have
:P: 5 :
N
N
X
X
ðAT Þk QAk : #
:ðAT Þk ::Q::Ak :
k50
N
X
k50
α2 λ2k
max :Q: 5
#
k50
α2
:Q:
1 2 λ2k
max
This contradicts the assumption and establishes necessity.
Uniqueness
Let P1 be a second solution of the Lyapunov equation. Then we can write it in the form of
infinite summation including Q, and then substitute for Q in terms of P using the Lyapunov
equation. This yields the equation
P1 5
N
N
X
X
ðAT Þk QAk 5 2
ðAT Þk ½AT PA 2 PAk
k50
52
k50
N
N
X
X
ðAT Þj PAj 1
ðAT Þk PAk 5 P
j51
k50
As in the nonlinear case, it is possible to relax the stability condition as
follows.
COROLLARY 11.2
The linear time-invariant discrete-time system of (11.40) is asymptotically stable if and only
if for any detectable pair (Ad, C), there exists a unique positive definite solution P to the discrete Lyapunov equation
ATd PAd 2 P 5 2C T C
(11.43)
PROOF
The proof follows the same steps as the theorem. We only show that zero values of the difference do not impact stability. We first obtain
11.4 Lyapunov stability theory
∆VðkÞ 5 Vðxðk 1 1ÞÞ 2 VðxðkÞÞ
5 xT ðkÞ½AT PA 2 PxðkÞ
5 2xT ðkÞCT CxðkÞ 5 2yT ðkÞyðkÞ # 0
If (Ad, C) is observable, y is zero only if x is zero. If the pair is only detectable, then the only
nonzero values of x for which y is zero are ones that correspond to stable dynamics and can
be ignored in our stability analysis.
Although using a semidefinite matrix in the stability test may simplify computation, checking the pair (Ad, C) for detectability (guaranteed by observability)
eliminates the gain from the simplification. However, the corollary is of theoretical interest and helps clarify the properties of the discrete Lyapunov equation.
The discrete Lyapunov equation is clearly a linear equation in the matrix P,
and by rearranging terms we can write it as a linear system of equations. The
equivalent linear system involves an n2 3 1 vector of unknown entries of P
obtained using the operation
stðPÞ 5 colfp1 ; p2 . . . ; pn g
P 5 p1 p2 . . . pn The n2 3 n2 coefficient matrix of the linear system is obtained using the
Kronecker product of matrices. In this operation, each entry of the first matrix is
replaced by the second matrix scaled by the original entry. The Kronecker product is thus defined by
A B 5 ½aij B
(11.44)
It can be shown that the Lyapunov equation (11.41) is equivalent to the linear
system
Lp 5 2q
p 5 stðPÞ
L 5 AT A T 2 I I
q 5 stðQÞ
(11.45)
Because the eigenvalues of any matrix are identical to those of its transpose,
the Lyapunov equation can be written in the form
Ad PATd 2 P 5 2Q
(11.46)
The first form of the Lyapunov equation of (11.41) is the controller form,
whereas the second form of (11.46) is known as the observer form.
11.4.5 MATLAB
To solve the discrete Lyapunov equation using MATLAB, we use the command
dlyap. The command solves the observer form of the Lyapunov equation.
cP 5 dlyapðA; QÞ
457
458
CHAPTER 11 Elements of Nonlinear Digital Control Systems
To solve the equivalent linear system, we use the Kronecker product and the
command
ckronðA0 ; A0 Þ
EXAMPLE 11.10
Use the Lyapunov approach with Q 5 I3 to determine the stability of the linear time-invariant system
2
3 2
32
3
20:2 20:2
0:4
x1 ðkÞ
x1 ðk 1 1Þ
4 x2 ðk 1 1Þ 5 5 4 0:5
0
1 54 x2 ðkÞ 5
x3 ðk 1 1Þ
x3 ðkÞ
0
20:4 20:5
Is it possible to investigate the stability of the system using Q 5 diag{0, 0, 1}?
Solution
Using MATLAB, we obtain the solution as follows:
cP 5 dlyapðA; QÞ % Observer form of the Lyapunov equation
1:5960
P 5 0:5666
0:0022
0:5666
3:0273
20:6621
0:0022
20:6621
1:6261
ceigðPÞ% Check the signs of the eigenvalues of P
1:1959
ans 5 1:6114
3:4421
Because the eigenvalues of P are all positive, the matrix is positive definite and the system
is asymptotically stable.
For Q 5 diag{0, 0, 1} 5 CTC, with C 5 [0, 0, 1], we check the observability of the pair
(A, C) using MATLAB and the rank test
crankðobsvða; ½0; 0; 1ÞÞ % obsv computes the observability matrix
ans 5 3
The observability matrix is full rank, and the system is observable. Thus, we can use the
matrix to check the stability of the system.
11.4.6 Lyapunov’s linearization method
It is possible to characterize the stability of an equilibrium for a nonlinear system
by examining its approximately linear behavior in the vicinity of the equilibrium.
Without loss of generality, we assume that the equilibrium is at the origin and
rewrite the state equation in the form
@f½xðkÞ 1 f 2 ½xðkÞ;
k 5 0; 1; 2; . . .
(11.47)
xðk 1 1Þ 5
@xðkÞ xðkÞ50
11.4 Lyapunov stability theory
where f2[.] is a function including all terms of order higher than one. We then
rewrite the equation in the form
xðk 1 1Þ 5 AxðkÞ 1 f 2 ½xðkÞ;
@f½xðkÞ A5
@xðkÞ k 5 0; 1; 2; . . .
(11.48)
xðkÞ50
In the vicinity of the origin the behavior is approximately the same as that of
the linear system
xðk 1 1Þ 5 AxðkÞ
(11.49)
This intuitive fact can be more rigorously justified using Lyapunov stability
theory, but we omit the proof. Thus, the equilibrium is stable if the linear approximation is stable and unstable if the linear approximation is unstable. If the linear
system (11.49) has an eigenvalue on the unit circle, then the stability of the nonlinear system cannot be determined from the first-order approximation alone. This
is because higher-order terms can make the nonlinear system either stable or
unstable. Based on our discussion, we have the following theorem.
THEOREM 11.6
The equilibrium point of the nonlinear system of (11.48) with linearized model (11.49) is as
follows:
•
•
•
Asymptotically stable if all the eigenvalues of A are inside the unit circle.
Unstable if one or more of the eigenvalues are outside the unit circle.
If A has one or more eigenvalues on the unit circle, then the stability of the nonlinear
system cannot be determined from the linear approximation.
EXAMPLE 11.11
Show that the origin is an unstable equilibrium for the system
x1 ðk 1 1Þ 5 2x1 ðkÞ 1 0:08x22 ðkÞ
x2 ðk 1 1Þ 5 x1 ðkÞ 1 0:1x2 ðkÞ 1 0:3x1 ðkÞx2 ðkÞ;
k 5 0; 1; 2; . . .
Solution
We first rewrite the state equations in the form
x1 ðk 1 1Þ
2 0
x1 ðkÞ
0:08x22 ðkÞ
;
5
1
x2 ðk 1 1Þ
1 0:1 x2 ðkÞ
0:3x1 ðkÞx2 ðkÞ
k 5 0; 1; 2; . . .
The state matrix A of the linear approximation has one eigenvalue 5 2 . 1. Hence, the
origin is an unstable equilibrium of the nonlinear system.
11.4.7 Instability theorems
The weakness of the Lyapunov approach for nonlinear stability investigation is
that it only provides sufficient stability conditions. Although no necessary and
459
460
CHAPTER 11 Elements of Nonlinear Digital Control Systems
sufficient conditions are available, it is possible to derive conditions for instability
and use them to test the system if one is unable to establish its stability. Clearly,
failure of both sufficient conditions is inconclusive, and it is only in the linear
case that we have the stronger necessary and sufficient condition.
THEOREM 11.7
The equilibrium point x 5 0 of the nonlinear discrete-time system
xðk 1 1Þ 5 f½xðkÞ;
k 5 0; 1; 2; . . .
(11.50)
is unstable if there exists a locally positive function for the system with locally uniformly positive definite changes along the trajectories of the system.
PROOF
For any motion along the trajectories of the system, we have
∆VðkÞ 5 Vðxðk 1 1ÞÞ 2 VðxðkÞÞ
5 Vðf½xðkÞÞ 2 Vðf½xðk 2 1ÞÞ . 0;
k 5 0; 1; 2; . . .
This implies that as the motion progresses, the value of V increases continuously. However,
because V is only zero with argument zero, the trajectories of the system will never converge
to the equilibrium at the origin and the equilibrium is unstable.
EXAMPLE 11.12
Show that the origin is an unstable equilibrium for the system
x1 ðk 1 1Þ 5 22x1 ðkÞ 1 0:08x22 ðkÞ
x2 ðk 1 1Þ 5 0:3x1 ðkÞx2 ðkÞ 1 2x2 ðkÞ;
k 5 0; 1; 2; . . .
Solution
Choose the Lyapunov function
VðxÞ 5 xT Px; P 5 diagfp1 ; 1g
The corresponding difference is given by
∆VðkÞ 5 p1 f½22x1 ðkÞ10:08x22 ðkÞ2 2 x21 ðkÞg 1 f½0:3x1 ðkÞx2 ðkÞ12x2 ðkÞ2 2 x22 ðkÞg
5 3p1 x21 ðkÞ 1 0:0064p1 x42 ðkÞ 1 f0:09x21 ðkÞ 1 ð1:2 2 0:32p1 Þx1 ðkÞ 1 3gx22 ðkÞ; k 5 0; 1; 2; . . .
pffiffiffi
We complete the squares for the last term by choosing p1 5 3:75ð1 2 3=2Þ 5 0:5024 and
reduce the difference to
pffiffiffi
∆VðkÞ 5 1:5072x21 ðkÞ 1 0:0032x42 ðkÞ 1 ð0:3x1 ðkÞ1 3Þ2 x22 ðkÞ
k 5 0; 1; 2; . . .
$ 1:5072x21 ðkÞ 1 0:0032x42 ðkÞ;
The inequality follows from the fact that the last term in ∆V is positive semidefinite. We conclude that ∆V is positive definite because it is greater than the sum of even powers and that
the equilibrium at x 5 0 is unstable.
11.4 Lyapunov stability theory
11.4.8 Estimation of the domain of attraction
We consider a system
xðk 1 1Þ 5 Ax 1 f½xðkÞ;
k 5 0; 1; 2; . . .
(11.51)
where the matrix A is stable and f(.) includes second-order terms and higher and
satisfies
2
:f½xðkÞ: # α:xðkÞ: ;
k 5 0; 1; 2; . . .
(11.52)
Because the linear approximation of the system is stable, we can solve the associated Lyapunov equation for a positive definite matrix P for any positive definite
matrix Q. This yields the Lyapunov function
VðxðkÞÞ 5 xT ðkÞPxðkÞ
and the difference
∆VðxðkÞÞ 5 xT ðkÞ½AT PA 2 PxðkÞ 1 2f T ½xðkÞPAxðkÞ 1 f T ½xðkÞPf½xðkÞ
To simplify this expression, we make use of the inequalities
2
2
λmin ðPÞ:x: # xT Px # λmax ðPÞ:x:
3
:f T PAx: # :f::PA::x: # α:PA::x:
(11.53)
where the norm jjAjj denotes the square root of the largest eigenvalue of the
matrix ATA or its largest singular value.
Using the inequality (11.53), and substituting from the Lyapunov equation, we
have
2
∆VðxðkÞÞ # ½α2 λmax ðPÞ:x: 1 2α:PA::x: 2 λmin ðQÞ:x:
2
(11.54)
Because the coefficient of the quadratic is positive, its second derivative is also
positive and it has negative values between its two roots. The positive root
defines a bound on jjxjj that guarantees a negative difference:
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
1
2
2λmax ðPAÞ 1 :PA: 1 αλmin ðQÞ
:x: ,
αλmax ðPÞ
An estimate of the domain of attraction is given by
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
1
2
2λmax ðPAÞ 1 :PA: 1 αλmin ðQÞ
BðxÞ 5 x: :x: ,
αλmax ðPÞ
EXAMPLE 11.13
Estimate the domain of attraction of the system
x1 ðk 1 1Þ
0:1
0
x1 ðkÞ
0:25x22 ðkÞ
5
1
2
x2 ðk 1 1Þ
21 0:5 x2 ðkÞ
0:1x1 ðkÞ
(11.55)
461
462
CHAPTER 11 Elements of Nonlinear Digital Control Systems
Solution
The nonlinear vector satisfies
:f½xðkÞ: # 0:5:xðkÞ:;
k 5 0; 1; 2; . . .
We solve the Lyapunov equation with Q 5 I2 to obtain
P5
2:4987
20:7018
20:7018
1:3333
whose largest eigenvalue is equal to 2.8281. The norm jjPAjj 5 1.8539 can be computed
with the MATLAB command
cnormðP AÞ
Our estimate of the domain of attraction is
8
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi)
<
1
pffiffiffiffiffiffiffiffiffiffiffiffiffiffi 21:8539 1 ð1:8539Þ2 1 0:25
BðxÞ 5 x: :x: ,
:
0:25 2:8281
5 fx: :x: , 0:0937g
x2(k)
3
2
1
0
–1
–2
–3
–2
–1.5
–1
–0.5
0
0.5
1
1.5
2 x1(k)
FIGURE 11.1
Phase portrait for the nonlinear system of Example 11.13.
The state portrait presented in Figure 11.1 shows that the estimate of the domain of
attraction is quite conservative and that the system is stable well outside the estimated
region.
11.5 Stability of analog systems with digital control
11.5 Stability of analog systems with digital control
Although most nonlinear design methodologies are analog, they are often digitally
implemented. The designer typically completes an analog controller design and
then obtains a discrete approximation of the analog controller. For example, a discrete approximation can be obtained using a differencing approach to approximate
the derivatives (see Chapter 6). The discrete-time approximation is then used with
the analog plant, assuming that the time response will be approximately the same
as that of the analog design. This section examines the question “Is the stability
of the digital control system guaranteed by the stability of the original analog
design?” As in the linear case discussed in Chapter 6, the resulting digital control
system may be unstable or may have unacceptable intersample oscillations.
We first examine the system of (11.1) with the digital control (11.2).
Substituting (11.2) in (11.1) gives the equation
x_ 5 fðxÞ 1 BðxÞuðkÞ 5 f k ðxÞ;
Let the solution of (11.56) be
tA½kT; ðk 1 1ÞTÞ
xðtÞ 5 gk ðxðkTÞ; tÞ
(11.56)
(11.57)
where gk is a continuous function of all its arguments. Then at the sampling
points, we have the discrete model
xðk 1 1Þ 5 gk ðxðkÞÞ;
k 5 0; 1; . . .
(11.58)
Theorem 11.8 shows that the stability of the discrete-time system of (11.58) is
a necessary condition for the stability of the analog system with digital control.
THEOREM 11.8: ANALOG SYSTEM WITH DIGITAL CONTROL
The continuous system of (11.56) with piecewise constant control is exponentially
stable only if the discrete-time system of (11.58) is exponentially stable.
PROOF
We show that the discrete-time system of (11.58) is stable for any stable system in (11.56).
For an exponentially stable continuous-time system, we have
:xðtÞ: # :xðkÞ:e2αk t ;
tA½kT; ðk 1 1ÞTÞ
(11.59)
for some positive constant αk. Hence, at the sampling points we have the inequality
:xðkÞ: # :xðk 2 1Þ:e2αk21 T ;
k 5 0; 1; . . .
Applying this inequality repeatedly gives
:xðkÞ: # :xð0Þ:e2αs T
# :xð0Þ:e2αm T ;
with αs 5
kP
21
k50
k 5 0; 1; . . .
αk and αm 5 mink αk . Thus, the discrete system is exponentially stable.
(11.60)
463
464
CHAPTER 11 Elements of Nonlinear Digital Control Systems
Theorem 11.8 only provides a necessary stability condition for the discretization of an analog plant with a digital controller. If a stable analog controller is
implemented digitally and used with the analog plant, its stability cannot be
guaranteed. The resulting system may be unstable, as Example 11.14 demonstrates.
EXAMPLE 11.14
Consider the system
_ 5 x2 ðtÞ 1 xðtÞuðtÞ
xðtÞ
with the continuous control
uðtÞ 5 2 xðtÞ 2 αx2 ðtÞ; α . 0
Then the closed-loop system is
_ 5 2 αx3 ðtÞ
xðtÞ
which is asymptotically stable.1
Now consider the digital implementation of the analog control
uðtÞ 5 2 xðkÞ 2 αx2 ðkÞ; α . 0;
tA½kT; ðk 1 1ÞTÞ
and the corresponding closed-loop system
_ 5 xðtÞfxðtÞ 2 ½αxðkÞ 1 1xðkÞg;
xðtÞ
tA½kT; ðk 1 1ÞTÞ
We solve the equation by separation of variables:
2
3
dxðtÞ
dxðtÞ 4 1
1
5
2 5 5 dt;
xðtÞðxðtÞ 2 βÞ
β
x2β
x
tA½kT; ðk 1 1ÞTÞ
β 5 ½αxðkÞ 1 1xðkÞ
Integrating from kT to (k 1 1)T, we obtain
#t5ðk11ÞT
#t5ðk11ÞT
β
5 dt
ln 12
xðtÞ
t5kT
12
t5kT
β
β
5 12
eT
xðk 1 1Þ
xðkÞ
xðk 1 1Þ 5
5
2
β
3
β 5 T
4
12 12
e
xðkÞ
½αxðkÞ 1 1xðkÞ
1 1 αxðkÞeT
For stability, we need the condition jx(k 1 1)j , jx(k)j—that is,
1 1 αxðkÞeT . 1 1 αxðkÞ
11.6 State plane analysis
This condition is satisfied for positive x(k) but not for all negative x(k)! For example, with
T 5 0.01 s and αx(k) 5 20.5, the LHS is 0.495 and the RHS is 0.5.
1
Slotine (1991, p. 66): x_1 cðxÞ 5 0 is asymptotically stable if c(.) is continuous and satisfies c(x)x . 0, ’x6¼0.
Here c(x) 5 x3.
11.6 State plane analysis
As shown in Section 11.4.6, the stability of an equilibrium of a nonlinear system
can often be determined from the linearized model of the system in the vicinity
of the equilibrium. Moreover, the behavior of system trajectories of linear discrete-time systems in the vicinity of an equilibrium point can be visualized for
second-order systems in the state plane. State plane trajectories can be plotted
based on the solutions of the state equations for discrete-time equations similar
to those for continuous-time systems (see Chapter 7). Consider the unforced
second-order difference equation
yðk 1 2Þ 1 a1 yðk 1 1Þ 1 a0 yðkÞ 5 0;
k 5 0; 1; . . .
(11.61)
The associated characteristic equation is
z2 1 a1 z 1 a0 5 ðz 2 λ1 Þðz 2 λ2 Þ 5 0
(11.62)
We can characterize the behavior of the system based on the location of the
characteristic values of the system λi, i 5 1, 2 in the complex plane. If the system is
represented in statespace form, then a similar characterization is possible using
the eigenvalues of the state matrix. Table 11.1 gives the names and characteristics
of equilibrium points based on the locations of the eigenvalues.
Trajectories corresponding to different types of equilibrium points are shown
in Figures 11.2 through 11.9. We do not include some special cases such as two
real eigenvalues equal to unity, in which case the system always remains at the
initial state. The reader is invited to explore other pairs of eigenvalues through
MATLAB simulation.
Table 11.1 Equilibrium Point Classification
Equilibrium Type
Eigenvalue Location
Stable node
Unstable node
Saddle point
Stable focus
Unstable focus
Vortex or center
Real positive inside unit circle
Real positive outside unit circle
Real eigenvalues with one inside and one outside unit circle
Complex conjugate or both real negative inside unit circle
Complex conjugate or both real negative outside unit circle
Complex conjugate on unit circle
465
466
CHAPTER 11 Elements of Nonlinear Digital Control Systems
y(k+1)
1
0.8
0.6
0.4
0.2
0
–0.2
–0.4
–0.6
–0.8
–1
–1 –0.8 –0.6 –0.4 –0.2
0
0.2 0.4 0.6 0.8
1 y(k)
0
1
5 y(k)
FIGURE 11.2
Stable node (eigenvalues 0.1, 0.3).
y(k+1)
10
8
6
4
2
0
–2
–4
–6
–8
–10
–5
–4
–3
FIGURE 11.3
Unstable node (eigenvalues 2, 3).
–2
–1
2
3
4
11.6 State plane analysis
y(k+1)
10
8
6
4
2
0
–2
–4
–6
–8
–10
–1.5
–1
–0.5
0
0.5
1
1.5 y(k)
FIGURE 11.4
Saddle point (eigenvalues 0.1, 3).
y(k+1)
1
0.8
0.6
0.4
0.2
0
–0.2
–0.4
–0.6
–0.8
–1
–1 –0.8 –0.6 –0.4 –0.2
FIGURE 11.5
Stable focus (eigenvalues 20.1, 20.3).
0
0.2 0.4 0.6 0.8
1 y(k)
467
468
CHAPTER 11 Elements of Nonlinear Digital Control Systems
y(k+1)
1
0.8
0.6
0.4
0.2
0
–0.2
–0.4
–0.6
–0.8
–1
–1 –0.8 –0.6 –0.4 –0.2
0
0.2 0.4 0.6 0.8
1 y(k)
FIGURE 11.6
Stable focus (eigenvalues 0.1 6 j0.3).
y(k+1)
× 1018
1
0.8
0.6
0.4
0.2
0
–0.2
–0.4
–0.6
–0.8
–1
–2
FIGURE 11.7
Unstable focus (eigenvalues 2 5, 23).
0
2
× 10
26 y(k)
11.6 State plane analysis
y(k+1)
15
10
5
0
–5
–10
–15
–20
–15
–10
–5
0
5
0
0.5
10
15
20 y(k)
FIGURE 11.8
Unstable focus (eigenvalues 0.1 6 j3).
y(k+1)
1.5
1
0.5
0
–0.5
–1
–1.5
–1.5
–1
–0.5
FIGURE 11.9
Center (eigenvalues cos(45 ) 6 j sin(45 )).
1
1.5 y(k)
469
470
CHAPTER 11 Elements of Nonlinear Digital Control Systems
11.7 Discrete-time nonlinear controller design
Nonlinear control system design for discrete-time systems is far more difficult
than linear design. It is also often more difficult than the design of analog nonlinear systems. For example, one of the most powerful approaches to nonlinear
system design is to select a control that results in a closed-loop system for which
a suitable Lyapunov function can be constructed. It is usually easier to construct a
Lyapunov function for an analog nonlinear system than it is for a discrete-time
system.
We discuss some simple approaches to nonlinear design for discrete-time
systems that are possible for special classes of systems.
11.7.1 Controller design using extended linearization
If one of the extended linearization approaches presented in Section 11.1 is applicable, then we can obtain a linear discrete-time model for the system and use it in
linear control system design. The nonlinear control can then be recovered from
the linear design. We demonstrate this approach with Example 11.15.
EXAMPLE 11.15
Consider the mechanical system
_ 1 cðxÞ 5 f
x€ 1 bðxÞ
with b(0) 5 0, c(0) 5 0. For the nonlinear damping bðxÞ
_ 5 0:1x_3 and the nonlinear spring
3
c(x) 5 0.1x 1 0.01x , design a digital controller for the system by redefining the input and
discretization. Use a sampling period T 5 0.02 s.
Solution
The state equations of the system are
x_1 5 x2
x_2 5 2bðx2 Þ 2 cðx1 Þ 1 f 5 u
Using the results for the discretization of the equivalent linear model as in Example 11.2,
we have
1 0:02
2 3 1024
uðkÞ
xðk 1 1Þ 5
xðkÞ 1
0
1
0:02
with x(k) 5 [x1(k) x2(k)]T 5 [x(k) x(k 1 1)]T.
We select the eigenvalues {0.1 6 j0.1} and design a state feedback controller for the
system as shown in Chapter 9. Using the MATLAB command place, we obtain the feedback
gain matrix
kT 5 2050 69:5 For a reference input r, we have the nonlinear control
f 5 u 1 bðx2 Þ 1 cðx1 Þ
uðkÞ 5 rðkÞ 2 kT xðkÞ
11.7 Discrete-time nonlinear controller design
r
Step
x1
Zero-Order
Hold
Out1
+
–
–
Subtract2
z2
Subsystem
1
s
x2
1
s
x1
Scope
Scope1
Zero-Order
Hold1
Polynomial1
P(u)
O(P) = 3
P(u)
O(P) = 2
Polynomial2
FIGURE 11.10
Simulation diagram for the system of Example 11.15.
1
r
P(u)
O(P) = 2
Polynomial2
2
x1
+
+
–
–
-KPolynomial1
P(u)
O(P) = 3
3
x2
Zero-Order
Hold
1
Out1
Subtract1
Gain1
+
+
Subtract2
-KGain
FIGURE 11.11
Controller simulation diagram of Example 11.15.
The simulation diagram for the system is shown in Figure 11.10, and the simulation
diagram for the controller block is shown in Figure 11.11. We select the amplitude of the
step input to obtain a steady-state value of unity using the equilibrium condition
xðk 1 1Þ 5
1
0
1
0:02
2 3 1024
ðr 2 ½ 2050 69:5 xðkÞÞ 5 xðkÞ 5
xðkÞ 1
0
1
0:02
This simplifies to
0:41
2 3 1024
r5
41
0:02
471
472
CHAPTER 11 Elements of Nonlinear Digital Control Systems
x1(k)
1.4
1.2
1
0.8
0.6
0.4
0.2
0
0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
1 t
FIGURE 11.12
Step response for the linear design of Example 11.15.
x2(k)
30
25
20
15
10
5
0
–5
–10
0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
1t
FIGURE 11.13
Velocity plot for the step response for the linear design of Example 11.15.
which gives the amplitude r 5 2050. The step response for the nonlinear system with digital
control in Figure 11.12 shows a fast response to a step input at t 5 0.2 s that quickly
settles to the desired steady-state value of unity. Figure 11.13 shows a plot of the velocity
for the same input.
11.7 Discrete-time nonlinear controller design
11.7.2 Controller design based on Lyapunov stability theory
Lyapunov stability theory provides a means of stabilizing unstable nonlinear systems
using feedback control. The idea is that if one can select a suitable Lyapunov function and force it to decrease along the trajectories of the system, the resulting system
will converge to its equilibrium. In addition, the control can be chosen to speed up
the rate of convergence to the origin by forcing the Lyapunov function to decrease
to zero faster.
To simplify our analysis, we consider systems of the form
xðk 1 1Þ 5 AxðkÞ 1 BðxðkÞÞuðkÞ
(11.63)
For simplicity, we assume that the eigenvalues of the matrix A are all inside
the unit circle. This model could approximately represent a nonlinear system in
the vicinity of its stable equilibrium.
THEOREM 11.9
The open-loop stable affine system with linear unforced dynamics and full-rank input matrix
for all nonzero x is asymptotically stable with the feedback control law
uðkÞ 5 2 ½BT ðxðkÞÞPBðxðkÞÞ21 BT ðxðkÞÞPAxðkÞ
(11.64)
where P is the solution of the discrete Lyapunov equation
AT PA 2 P 5 2Q
and Q is an arbitrary positive definite matrix.
(11.65)
PROOF
For the Lyapunov function
VðxðkÞÞ 5 xT ðkÞPxðkÞ
the difference is given by
∆VðkÞ 5 Vðxðk 1 1ÞÞ 2 VðxðkÞÞ
5 xT ðkÞ½AT PA 2 PxðkÞ 1 2uT ðkÞBT PAx 1 uT ðkÞBT PBuðkÞ
where the argument of B is suppressed for brevity. We minimize the function with respect to
the control to obtain
@∆VðkÞ
5 2 BT PAx 1 BT PBuðkÞ 5 0
@uðkÞ
which we solve for the feedback control law using the full-rank condition for B.
By the assumption of open-loop stability, we have
∆VðkÞ 5 2xT ðkÞQxðkÞ 1 xT ðkÞAT PBuðkÞ;
Q.0
We substitute for the control and rewrite the equation as
∆VðkÞ 5 2xT ðkÞfQ 1 AT PB½BT PB21 BT AgxðkÞ
5 2xT ðkÞfQ 1 AT MAgxðkÞ
M5 PB½BT PB21 BTP
473
474
CHAPTER 11 Elements of Nonlinear Digital Control Systems
Because the matrix P is positive definite and B is full rank, the matrix BTPB, and consequently its inverse, must be positive definite. Thus, the matrix M is positive semidefinite,
and the term 2 xT(k)ATMAx(k) is negative semidefinite. We conclude that the difference of
the Lyapunov function is negative definite, because it is the sum of a negative definite
term and a negative semidefinite term, and that the closed-loop system is asymptotically
stable.
EXAMPLE 11.16
Design a controller to stabilize the origin for the system
x1 ðk 1 1Þ 5 0:2x1 ðkÞ 1 x2 ðkÞuðkÞ
x2 ðk 1 1Þ 5 0:2x2 ðkÞ 1 ½1 1 0:4x1 ðkÞuðkÞ;
k 5 0; 1; 2; . . .
Solution
We rewrite the system dynamics in the form
0:2 0
x1 ð k Þ
x2 ðkÞ
x1 ðk 1 1Þ
5
1
uðkÞ
0 0:2 x2 ðkÞ
x2 ðk 1 1Þ
1 1 0:4x1 ðkÞ
The state matrix is diagonal with two eigenvalues equal to 0.2 , 1. Hence, the system is
stable. We choose the Lyapunov function
VðxÞ 5 xT Px; P 5 diagfp1 ; 1g; p1 .0
for which Q 5 0.96P is positive definite.
The corresponding stabilizing control is given by
uðkÞ 5 2 ½BT ðxðkÞÞPBðxðkÞÞ21 BT ðxðkÞÞPAxðkÞ
52
0:2
½p1 x2 ðkÞ
ð110:4x1 ðkÞÞ2 1 p1 x22 ðkÞ
1 1 0:4x1 ðkÞxðkÞ
We choose p1 5 5 and obtain the stabilizing control
uðkÞ 5 2
1
x2 ðkÞ 0:2 1 0:08x1 ðkÞ xðkÞ
ð110:4x1 ðkÞÞ2 1 5x22 ðkÞ
11.8 Input-output stability and the small gain theorem
Consider a nonlinear dynamic system as an operation that maps a sequence of
m 3 1 input vectors u(k) to a sequence of l 3 1 output vector y(k). The equation
representing this mapping is
y 5 NðuÞ
(11.66)
where N(.) represents the operation performed by the dynamic system,
u 5 col{u(k), k 5 0, 1, . . ., N} represents the input sequence, and y 5 col{y(k),
k 5 0, 1, . . . , N} represents the output sequence. We restrict our analysis of nonlinear
systems to the class of causal systems where the output at any given instant is
11.8
Input-output stability and the small gain theorem
caused by the history of system inputs and is unaffected by future inputs. Similarly
to Chapter 4, we define input-output stability based on the magnitude of the output
vector over its history for a bounded input vector. The magnitude of a vector over
its history is defined using the following norm:
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ffi
XN
2
(11.67)
:uðkÞ:
:u:s 5
k50
The norm of each vector in the summation can be any vector norm for the
m 3 1 vectors u(k), k 5 0, 1, . . ., N, but we restrict our discussion to the 2-norm.
Thus, the norm of (11.67) is the 2-norm of an infinite-dimensional vector
obtained by stacking the vectors u(k), k 5 0, 1, . . ., N. We can define input
output stability for nonlinear systems as in Chapter 4.
DEFINITION 11.5: INPUT-OUTPUT STABILITY
A system is input-output stable if for any bounded input sequence
:u:s , Ku
(11.68)
the output sequence is also bounded—that is,
:y:s , Ky
(11.69)
For a system to be input-output stable, the system must have an upper bound on the ratio of
the norm of its output to that of its input. We call this upper bound the gain of the nonlinear
system and define it as
γ 5 max
:u:s
:y:s
(11.70)
:u:s
EXAMPLE 11.17
Find the gain of a saturation nonlinearity (Figure 11.14) governed by
8
< Ku; juj , L
y 5 NðuÞ 5
KL; u $ L
:
2KL; u # 2 L
K . 0; L . 0
Solution
The output is bounded by
y 5
K juj; juj , L
KL; juj $ L
The norm of the output is
2
:y:s 5
N
X
jyðkÞj2 # K
k50
Thus, the gain of the system is the constant γsat 5 K.
N
X
k50
juðkÞj2
475
476
CHAPTER 11 Elements of Nonlinear Digital Control Systems
y
KL
−L
u
L
− KL
FIGURE 11.14
Saturation nonlinearity.
EXAMPLE 11.18
Find the gain of a causal linear system with impulse response
GðkÞ; k 5 0; 1; 2; . . .
Solution
The response of the system to any input u(k), k 5 0, 1, 2, . . . is given by the convolution
summation
yðkÞ 5
N
X
GðiÞuðk 2 iÞ; k 5 0; 1; 2; . . .
i50
The square of the norm of the output sequence is
2
:y:s 5
N
X
2
:yðkÞ: #
k50
N
N
X
X
2
:
GðiÞuðk2iÞ:
k50
i50
By the triangle inequality for norms, we have
2
:y:s #
N X
N
X
2
2
:GðiÞ: :uðk2iÞ:
k50 i50
N
N
X
X
2
2
2
2
5
:GðiÞ:
:uðk2iÞ: 5 :G:s :u:s
i50
k50
Thus, the gain of the linear causal system is the norm of its impulse response sequence
γ G 5 :G:s 5
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
N
X
2
:GðiÞ:
i50
Recall from Theorem 8.2 that a linear system is said to be BIBO stable if its impulse
response sequence is absolutely summable, as guaranteed if the system gain defined here
is finite.
11.8
Input-output stability and the small gain theorem
We can also express the gain of a linear system in terms of its frequency
response as follows.
THEOREM 11.10
The gain of a stable linear system is given by
γ G 5 :G:s 5 max :GðejωT Þ:
(11.71)
ω
PROOF
Parseval’s identity states that the energy of a signal in the time domain is equal to its energy
in the frequency domain—that is,
2
:y:s 5
XN
2
:yðkÞ: 5
k50
1
2π
ðπ
2
2π
:yðejωT Þ: dω
(11.72)
For a stable system, the energy of the output signal is finite, and we can use Parseval’s
identity to write the norm as
:y:s 5
1
2π
5
1
2π
2
#
ðπ
2π
ðπ
2π
1
2π
:yðe2jωT ÞT yðejωT Þ:dω
:uðe2jωT ÞT Gðe2jωT ÞT GðejωT ÞuðejωT Þ:dω
ðπ
:Gðe2jωT ÞT GðejωT Þ::uðejωT Þ: dω
2
2π
2
# maxω :GðejωT Þ: 3
1
2π
ðπ
2π
2
:uðejωT Þ: dω
Using Parseval’s identity on the input terms, we have
2
2
2
:y:s # max :GðejωT Þ: :u:s
ω
Hence, the gain of the system is given by (11.71).
EXAMPLE 11.19
Verify that the expression of (11.71) can be used to find the gain of the linear system with
transfer function
GðzÞ 5
z
;
z2a
jaj , 1
Solution
The impulse response of the linear system is
gðkÞ 5 ak ;
k 5 0; 1; 2; . . .; jaj , 1
477
478
CHAPTER 11 Elements of Nonlinear Digital Control Systems
and its gain is
2
:g:s 5
N N
X
X
gðkÞ2 5
jaj2k 5
k50
k50
1
1 2 jaj2
The frequency response of the system is
GðejωT Þ 5
ejωT
1
5
ejωT 2 a 1 2 ae2jωT
The maximum magnitude is attained when the denominator is minimized to obtain
2
1
max :GðejωT Þ: 5 max ω
ω 12ae2jωT 2
2
1
5 max ω 12acosðωTÞ1jasinðωTÞ
2
1
1
5
5 max ω 122acosðωTÞ1a2 1 2 jaj2
We are particularly interested in the stability of closed-loop systems of
the form in Figure 11.15, where two causal nonlinear systems are connected
in a closed-loop configuration. The stability of the system is governed by
Theorem 11.11.
THEOREM 11.11: THE SMALL GAIN THEOREM
The closed-loop system in Figure 11.15 is input-output stable if the product of the gains
of all the systems in the loop is less than unity—that is,
2
Li51 γNi , 1
r
e
(11.73)
y
N1(.)
N2(.)
FIGURE 11.15
Block diagram of a nonlinear closed-loop system.
11.8
Input-output stability and the small gain theorem
PROOF
The output of the system is given by
y 5 N1 ðeÞ
and the error is given by
e 5 r 2 N2 ðyÞ
Using norm inequalities, we have
:e: # :r: 1 γ N2 :y: # :r: 1 γN2 γ N1 :e:
Solving for the norm of the error, we obtain
:e: #
:r:
1 2 γ N2 γN1
which is finite if (11.73) is satisfied.
Note that Theorem 11.11 is applicable if the forward path, the feedback path,
or both are replaced by a cascade of nonlinear systems so that the condition
(11.73) can be generalized to n nonlinear systems in a feedback loop as
n
Li51 γ Ni , 1
(11.74)
The results of the small gain theorem provide a conservative stability condition when applied to linear systems as compared to the Nyquist criterion. For a
single-input-single-output system, the theorem implies that a closed-loop system
is stable if its loop gain has magnitude less than unity. This corresponds to openloop stable systems whose frequency response plots are completely inside the
gray unit circle shown in Figure 11.16. These frequency response plots clearly do
not encircle the 21 point and are obviously stable using the Nyquist criterion.
However, the Nyquist criterion uses phase data and can show the stability for a
wider class of systems, including ones for which the circle criterion fails.
Im [G(e jωT)]
•
–1
FIGURE 11.16
Unit circle in the complex plane.
Re [G(e jωT)]
479
480
CHAPTER 11 Elements of Nonlinear Digital Control Systems
EXAMPLE 11.20
Simulate a feedback loop with linear subsystem transfer function (T 5 0.01 s)
GðzÞ 5
z 1 0:5
z2 1 0:5z 1 0:3
in series with an amplifier of variable gain and a symmetric saturation nonlinearity with
slope and saturation level unity. Investigate the stability of the system and discuss your
results referring to the small gain theorem for two amplifier gains: 1 and 0.2.
Solution
We use SIMULINK to simulate the system, and the simulation diagram is shown in
Figure 11.17.
The maximum magnitude of the frequency response of the linear subsystem can be
obtained using the command
c½mag; ph 5 bodeðgÞ
The maximum magnitude is approximately 5.126. By the small gain theorem, the stability condition for the gain of the nonlinearity is γN , 1/5.125 0.195.
1. The gain of the nonlinear block with slope and saturation level unity is also unity. Thus,
by the small gain theorem, the stability of the system cannot be guaranteed with an
amplifier gain of unity. This is verified by the simulation results of Figure 11.18.
1.z+.50
0.2
2
z +0.5z+0.8
Gain
Saturation
Discrete
Transfer Fcn
Scope
FIGURE 11.17
Simulation diagram of the closed-loop system with saturation nonlinearity.
4
2
0
–2
–4
0
0.5
1
FIGURE 11.18
Response of the closed-loop system with initial conditions [1, 0] and unity gain.
1.5
11.8
Input-output stability and the small gain theorem
0.6
0.4
0.2
0
–0.2
–0.4
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
FIGURE 11.19
Response of the closed-loop system with initial conditions [1, 0] and 0.2 gain.
However, if the amplifier gain is reduced to 0.1, then the overall gain of the linear subsystem is reduced to about 0.513, which is also equal to the product of the gains in
the loop. By the small gain theorem, the stability of the system is guaranteed.
2. The results of the small gain theorem are conservative, and the system may be
stable for larger gain values. If fact, the simulation results of Figure 11.19 show that
the system is stable for a gain of 0.2, for which the small gain theorem fails.
11.8.1 Absolute stability
In this section, we consider the stability of the SISO system of Figure 11.20 for a
class of nonlinearities N(.). We obtain sufficient conditions for the stability of the
closed-loop system if the nonlinearity belongs to the following class (Definition 11.6).
r
y
e
G(z)
N(.)
FIGURE 11.20
Block diagram of closed-loop system with linear and nonlinear blocks.
DEFINITION 11.6
Sector Bound Nonlinearity
A nonlinearity (Figure 11.21) N(.) belongs to the sector [kl, ku] if it satisfies the condition
kl #
NðyÞ
# ku
y
481
482
CHAPTER 11 Elements of Nonlinear Digital Control Systems
Ku
N(y)
Kl
D(Kl , Ku)
y
–1/Kl
–1/Ku
FIGURE 11.21
Sector bound nonlinearity and the associated disc.
The stability of the loop for any sector bound nonlinearity is defined as follows.
We associate the sector bound nonlinearity with the disk D(kl, ku) bounded by a
circle with its center on the negative real axis with endpoints 21/ku and 21/kl.
DEFINITION 11.7
Absolute Stability
A nonlinear feedback system of the form in Figure 11.20 is absolutely stable with respect to
the sector [kl, ku] if the origin of the state space x 5 0 is globally asymptotically stable for all
nonlinearities belonging to the sector [kl, ku].
Theorem 11.12 gives sufficient conditions for the absolute stability of the closed
system with any stable linear block.
THEOREM 11.12: THE CIRCLE CRITERION
The closed-loop system of Figure 11.20 is absolutely stable with G(z) stable and nonlinearity
N(.) belonging to sector [kl, ku] if one of the following conditions is satisfied:
a.
b.
c.
d.
e.
a.
If 0 , kl , ku and the Nyquist plot of G(ejωT) does not enter or encircle the disc D(kl, ku).
If 0 5 kl , ku and the Nyquist plot of G(ejωT) lies to the right of the line z 5 21/ku.
If kl , 0 , ku and the Nyquist plot of G(ejωT) lies in the interior of the disc D(kl, ku).
If kl , ku 5 0 and the Nyquist plot of G(ejωT) lies to the left of the line z 5 21/kl.
If kl , ku , 0 and the Nyquist plot of G(ejωT) does not enter or encircle the disc D(kl, ku). ’
0 , k l , ku
ku
N(y)
kl
D (kl, ku)
y
–1/kl
–1/ku
11.8
Input-output stability and the small gain theorem
b. 0 5 kl , ku
ku
N(y)
kl
y
z = –1/ku
c. kl , 0 , ku
N(y)
ku
D (kl, ku)
a.
kl
d. kl , ku 5 0
N(y)
ku = 0
kl
z = –1/kl
e. kl , ku , 0
N(y)
y
D (kl, ku)
ku
kl
–1/kl
–1/ku
483
CHAPTER 11 Elements of Nonlinear Digital Control Systems
The circle criterion provides a sufficient stability condition, but its results can
be conservative because it is only based on bounds on the nonlinearity without
considering its shape. An overoptimistic estimate of the stability sector can be
obtained using the Jury criterion with the nonlinearity replaced by a constant gain
as in Section 4.5. The truth will typically lie somewhere between these two
estimates, as seen from Example 11.21.
EXAMPLE 11.21
For the furnace and actuator in Example 4.10, determine the stability sector for the
system and compare it to the stable range for a linear gain block.
Solution
The transfer function of the actuator and furnace is
Ga ðzÞGZAS ðzÞ 5 1025
4:711z 1 4:664
z3 2 2:875z2 1 2:753z 2 0:8781
and the corresponding Nyquist plot is shown in Figure 11.22. We use the circle
criterion to examine the stability of the closed-loop system with a sector bound
nonlinearity.
The plot lies inside a circle of radius 10 1 A, which corresponds to a sector (20.1, 0.1).
It is to the right of the line z 5 20.935, which corresponds to a sector (0, 1.05). Note that
we cannot conclude that the system is stable in the sector (20.1, 1.05); that is, we cannot
combine stability sectors. Nevertheless, because the circle criterion gives conservative
results, the actual stable sector may be wider than our estimates indicate. The Jury criterion
gives the stable range (2 0.099559, 3.47398). The lower bound is approximately the same
as predicted by the circle criterion, but the upper bound is more optimistic. Simulation
Nyquist Diagram
6
4
Imaginary Axis
484
2
0
–2
System: l
Real: -0.935
Imag: 1.4
Frequency (rad/sec): -1.8
–4
–6
–2
0
2
FIGURE 11.22
Nyquist plot of furnace and actuator.
4
Real Axis
6
8
10
Problems
2.5
2
1.5
1
0.5
0
0
1
2
3
4
5
6
7
8
9
10
FIGURE 11.23
Simulation results for the nonlinear system for a saturation nonlinearity with a linear
gain of unity and saturation levels (21, 1).
results, shown in Figure 11.23, verify the stability of the system with saturation nonlinearity
in the sector [21, 1]. This is a much larger sector than the conservative estimate of the
circle criterion but is well inside the range obtained using the Jury criterion.
Resources
Apostol, T.M., 1975. Mathematical Analysis. Addison-Wesley, Reading, MA.
Fadali, M.S., 1987. Continuous drug delivery system design using nonlinear decoupling:
A tutorial. IEEE Trans. Biomedical Engineering 34 (8), 650653.
Goldberg, S., 1986. Introduction to Difference Equations. Dover, Mineola, NY.
Kalman, R.E., Bertram, J.E., 1960. Control system analysis and design via the “Second Method”
of Lyapunov II: Discrete time systems. J. Basic Engineering Trans. ASME 82 (2), 394400.
Khalil, H.K., 2002. Nonlinear Systems. Prentice Hall, Upper Saddle River, NJ.
Kuo, B.C., 1992. Digital Control Systems. Saunders, Fort Worth, TX.
LaSalle, J.P., 1986. The Stability and Control of Discrete Processes. Springer-Verlag, New York.
Mickens, R.E., 1987. Difference Equations. Van Nostrand Reinhold, New York.
Mutoh, Y., Shen, T., Nikiforuk, P.N., 1996. Absolute stability of Discrete Nonlinear
Systems, Techniques in Discrete and Continuous Robust Systems, Control and
Dynamic Systems, Vol. 74. Academic Press, pp. 225252.
Oppenheim, A.V., Willsky, A.S., Nawab, S.H., 1996. Signals and Systems. Prentice Hall,
Upper Saddle River, NJ.
Rugh, W.J., 1996. Linear System Theory. Prentice Hall, Upper Saddle River, NJ.
Slotine, J.-J., 1991. Applied Nonlinear Control. Prentice Hall, Englewood Cliffs, NJ.
PROBLEMS
11.1
Discretize the following system:
x_1
23x1 2 x21 =ð2x22 Þ
x21
uðtÞ
5
1
x_2
x22 =x1 1 1
0
485
486
CHAPTER 11 Elements of Nonlinear Digital Control Systems
11.2
The equations for rotational maneuvering2 of a helicopter are given by
0
1
0
1
1 @Iz 2 Iy A _
I
2
I
z
y
Aψ_ θ_
ψ sinð2θÞ 2 mgA cosðθÞ 1 2@
θ€ 5 2 2
Ix
Ix
sinð2θÞ
Tp
ðIz 1 Iy Þ 1 ðIz 2 Iy Þcosð2θÞ
ψ_ 5
1
Ty
ðIz 1 Iy Þ 1 ðIz 2 Iy Þcosð2θÞ
where
Ix, Iy, and Iz 5 moments of inertia about the center of gravity
m 5 total mass of the system
g 5 acceleration due to gravity
θ and ψ 5 pitch and yaw angles in radians
Tp and Ty 5 pitch and yaw input torques
Obtain an equivalent linear discrete-time model for the system, and derive
the equations for the torque in terms of the linear system inputs.
11.3
A single-link manipulator3 with a flexible link has the equation of
motion
Iθ€ 1 MgL sinðθÞ 2 mgA cosðθÞ 1 kðθ 2 ψÞ
50
Jψ€ 1 kðψ 2 θÞ 5 τ
where
L 5 distance from the shaft to the center of gravity of the link
M 5 mass of the link
I 5 moment of inertia of the link
J 5 moment of inertia of the joint
K 5 rotational spring constant for the flexible joint
L 5 distance between the center of gravity of the link and the flexible
joint
θ and ψ 5 link and joint rotational angles in radians
τ 5 applied torque
Obtain a discrete-time model of the manipulator (Figure P11.1).
2
Elshafei, A.L., Karray, F., 2005. Variable structure-based fuzzy logic identification of a class of
nonlinear systems. IEEE Trans. Control Systems Tech. 13 (4), 646653.
3
Spong, M.W., Vidyasagar, M., 1989. Robot Dynamics and Control. Wiley, New York, pp. 269273.
Problems
M
c.g.
L
k
ψ, τ
θ
FIGURE P11.1
Schematic of a single-link manipulator.
11.4
Solve the nonlinear difference equation
½yðk 1 2Þ½yðk11Þ22 ½yðkÞ1:25 5 uðkÞ
with zero initial conditions and the input u(k) 5 e2k.
11.5
Determine the equilibrium point for the system
x1 ðk 1 1Þ
0
2x1 ðkÞ=9 1 2x22 ðkÞ
5
uðkÞ
1
x2 ðk 1 1Þ
x1 ðkÞ
2x2 ðkÞ=9 1 0:4x21 ðkÞ
(a) Unforced
(b) With a fixed input ue 5 1
11.6
Use the Lyapunov approach to show that if the function f(x) is a contraction, then the system x(k 1 1) 5 f[x(k)] is asymptotically stable.
11.7
Obtain a general expression for the eigenvalues of a 2 3 2 matrix, and
use it to characterize the equilibrium points of the second-order system
with the given state matrix:
0:9997
0:0098
(a)
20:0585 0:9509
3 1
(b)
1 2
0:3 20:1
(c)
0:1
0:2
1:2 20:4
(d)
0:4
0:8
11.8
Determine the stability of the origin using the linear approximation for
the system
x1 ðk 1 1Þ5 0:2x1 ðkÞ 1 1:1x32 ðkÞ
x2 ðk 1 1Þ 5 x1 ðkÞ 1 0:1x2 ðkÞ 1 2x1 ðkÞx22 ðkÞ;
k 5 0; 1; 2; . . .
487
488
CHAPTER 11 Elements of Nonlinear Digital Control Systems
11.9
Verify the stability of the origin using the Lyapunov approach, and estimate the rate of convergence to the equilibrium
x1 ðk 1 1Þ5 0:1x1 ðkÞx2 ðkÞ 2 0:05x22 ðkÞ
x2 ðk 1 1Þ 5 2 0:5x1 ðkÞx2 ðkÞ 1 0:05x32 ðkÞ;
k 5 0; 1; 2; . . .
11.10
Show that the convergence of the trajectories of a nonlinear discrete-time
system x(k 1 1) 5 f[x(k)] to a known nominal trajectory x (k) is equivalent
to the stability of the dynamics of the tracking error e(k) 5 x(k) 2 x (k).
11.11
Prove that the scalar system
xðk 1 1Þ 5 2ax3 ðkÞ;
a.0
pffiffiffi
is locally asymptotically stable in the region xðkÞ # 1= a.
11.12
Use Lyapunov stability theory to investigate the stability of the system
x1 ðk 1 1Þ 5
ax1 ðkÞ
a 1 bx22 ðkÞ
x2 ðk 1 1Þ 5
bx2 ðkÞ
;
b 1 ax21 ðkÞ
a . 0; b . 0
11.13
Use the Lyapunov approach to determine the stability of the discrete-time
linear time-invariant systems
0:3 20:1
(a)
0:1 0:22
2
3
0:3 20:1 0
(b) 4 0:1 0:22 0:2 5
0:4
0:2
0:1
11.14
Show that the origin is an unstable equilibrium for the system
x1 ðk 1 1Þ 5 21:4x1 ðkÞ 1 0:1x22 ðkÞ
x2 ðk 1 1Þ 5 1:5x2 ðkÞðÞð0:1x1 ðkÞ 1 1Þ;
k 5 0; 1; 2; . . .
11.15
Estimate the domain of attraction of the system
x1 ðk 1 1Þ
0:2
0:3 x1 ðkÞ
0:3x22 ðkÞ
5
1
20:4 0:5 x2 ðkÞ
x2 ðk 1 1Þ
0:36x21 ðkÞ
11.16
Design a controller to stabilize the origin for the system
x1 ðk 1 1Þ5 0:4x1 ðkÞ 1 0:5x2 ðkÞ 1 x22 ðkÞuðkÞ
x2 ðk 1 1Þ 5 0:1x1 ðkÞ 1 0:2x2 ðkÞ 1 ½x2 ðkÞ 1 x1 ðkÞuðkÞ;
k 5 0; 1; 2; . . .
Computer Exercises
COMPUTER EXERCISES
11.17
Write a MATLAB program to generate phase plane plots for a discretetime second-order linear time-invariant system. The function should
accept the eigenvalues of the state matrix and the initial conditions
needed to generate the plots.
11.18
Design a controller for the nonlinear mechanical system described in
_ 5 0:25x_5 , the nonlinear
Example 11.15 with the nonlinear damping bðxÞ
3
spring c(x) 5 0.5x 1 0.02 3 , T 5 0.02 s, and the desired eigenvalues for
the linear design equal to {0.2 6 j0.1}. Determine the value of the reference input for a steady-state position of unity, and simulate the system
using Simulink.
11.19
Design a stabilizing digital controller with a sampling period T 5 0.01 s
for a single-link manipulator using extended linearization, then simulate
the system with your design. The equation of motion of the manipulator
is given by θ€ 1 0:01 sinðθÞ 1 0:01θ 1 0:001θ3 5 τ.
Assign the eigenvalues of the discrete-time linear system to
{0.6 6 j0.3}.
Hint: Use Simulink for your simulation, and use a ZOH block, or a
discrete filter block with both the numerator and denominator set to 1
for sampling.
11.20
Some chemical processes, such as a distillation column, can be modeled
as a series of first-order processes with the same time constant. For the
process with transfer function
GðsÞ 5
1
ðs11Þ10
design a digital proportional controller with a sampling period T 5 0.1 by
applying the tangent method and the Ziegler-Nichols rules (see Sections
5.5 and 6.3.4). Use the circle criterion to show that the closed-loop system with actuator saturation nonlinearity of slope unity is asymptotically
stable. Obtain the step response of the closed-loop system to verify the
stability of the closed-loop system with actuator saturation.
489
CHAPTER
Practical Issues
12
OBJECTIVES
After completing this chapter, the reader will be able to do the following:
1. Write pseudocode to implement a digital controller.
2. Select the sampling frequency in the presence of antialiasing filters and
quantization errors.
3. Implement a PID controller effectively.
4. Design a controller that addresses changes in the sampling rate during control
operation.
5. Design a controller with faster input sampling than output sampling.
Successful practical implementation of digital controllers requires careful attention
to several hardware and software requirements. In this chapter, we discuss the
most important of these requirements and their influence on controller design.
We analyze the choice of the sampling frequency in more detail (already discussed
in Section 2.9) in the presence of antialiasing filters and the effects of quantization,
rounding, and truncation errors. In particular, we examine the effective implementation of a proportionalintegralderivative (PID) controller. Finally, we examine
changing the sampling rate during control operation as well as output sampling at
a slower rate than that of the controller.
12.1 Design of the hardware and software architecture
The designer of a digital control system must be mindful of the fact that the control algorithm is implemented as a software program that forms part of the control
loop. This introduces factors that are not present in analog control loops. This section discusses several of these factors.
12.1.1 Software requirements
During the design phase, designers make several simplifying assumptions that
affect the implemented controller. They usually assume uniform sampling with
negligible delay due to the computation of the control variable. Thus, they assume
no delay between the sampling instant and the instant at which the computed
491
492
CHAPTER 12 Practical Issues
control value is applied to the actuator. This instantaneous execution assumption
is not realistic because the control algorithm requires time to compute its output
(Figure 12.1). If the computational time is known and constant, we can use the
modified z-transform (see Section 2.7) to obtain a more precise discrete model.
However, the computational time of the control algorithm can vary from one sampling period to the next. The variation in the computational delay is called the
control jitter. For example, control jitter is present when the controller implementation utilizes a switching mechanism.
Digital control systems have additional requirements such as data storage and
user interface, and their proper operation depends not only on the correctness of
their calculations but also on the time at which their results are available. Each
task must satisfy either a start or completion timing constraint. In other words, a
digital control system is a real-time system. To implement a real-time system,
we need a real-time operating system that can provide capabilities such as multitasking, scheduling, and intertask communication, among others. In a multitasking
environment, the value of the control variable must be computed and applied over
each sampling interval regardless of other tasks necessary for the operations of
the overall control system. Hence, the highest priority is assigned to the computation and application of the control variable.
Clearly, the implementation of a digital control system requires skills in software engineering and computer programming. There are well-known programming guidelines that help minimize execution time and control jitter for the
control algorithm. For example, if-then-else and case statements must be avoided
y
t/T
u
Te
T
k−1
k
t/T
k+1
FIGURE 12.1
Execution time Te for the computation of the control variable for a system with sampling
period T.
12.1 Design of the hardware and software architecture
as much as possible because they can lead to paths of different lengths and, consequently, paths with different execution times. The states of the control variable
must be updated after the application of the control variable. Finally, the software
must be tested to ensure that no errors occur. This is known as software verification. In particular, the execution time and the control jitter must be measured to
verify that they can be neglected relative to the sampling period, and memory
usage must be analyzed to verify that it does not exceed the available memory.
Fortunately, software tools are available to make such analysis possible.
EXAMPLE 12.1
Write pseudocode that implements the following controller:
CðzÞ 5
UðzÞ 10:5z 2 9:5
5
EðzÞ
z21
Then propose a possible solution to minimize the execution time (see Figure 12.1).
Solution
The difference equation corresponding to the controller transfer function is
uðkÞ 5 uðk 2 1Þ 1 10:5eðkÞ 2 9:5eðk 2 1Þ
This control law can be implemented by writing the following code:
function controller
% This function is executed during each sampling period
% r is the value of the reference signal
% u1 and e1 are the values of the control variable and of the control
% error respectively for the previous sampling period
y 5 read_ADC(ch0) % Read the process output from channel 0 of the ADC
e 5 r 2 y; % Compute the tracking error
u 5 u1 1 10.5 e 2 9.5 e1; % Compute the control variable
u1 5 u; % Update the control variable for the next sampling period
e1 5 e; % Update the tracking error for the next sampling period
write_DAC(ch0,u); % Output the control variable to channel 0 of the
%DAC
To decrease the execution time, two tasks are assigned different priorities (using a
real-time operating system):
Task 1 (Maximum Priority)
y 5 read_ADC(ch0) % Read the process output from channel 0 of the ADC
e 5 r 2 y; % Compute the tracking error
u 5 u1 1 10.5 e 2 9.5 e1; % Compute the control variable
write_DAC(ch0,u); % Write the control variable to channel 0 of the
%DAC
Task 2
u1 5 u; % Update the control variable for the next sampling period
e1 5 e; % Update the tracking error for the next sampling period
493
494
CHAPTER 12 Practical Issues
12.1.2 Selection of ADC and DAC
The ADC and DAC must be sufficiently fast for negligible conversion time relative to the sampling period. In particular, the conversion delay of the ADC creates
a negative phase shift, which affects the phase margin of the system and must be
minimized to preserve the stability margins of the system. In addition, the word
length of the ADC affects its conversion time. With the conversion time provided
by standard modern analog-to-digital converters, this is not a significant issue in
most applications.
The choice of the ADC and DAC word length is therefore mainly determined
by the quantization effects. Typically, commercial ADCs and DACs are available in the range of 8 to 16 bits. An 8-bit ADC provides a resolution of a 1 in 28,
which corresponds to an error of 0.4%, whereas a 16-bit ADC gives an error of
0.0015%.
Clearly, the smaller the ADC resolution, the better the performance, and therefore a 16-bit ADC is preferred. However, the cost of the component increases as
the word length increases, and the presence of noise might render the presence of
a high number of bits useless in practical applications. For example, if the sensor
has a 5 mV noise and a 5 V range, there is no point in employing an ADC with
more than 10 bits because its resolution of 1 in 210 corresponds to an error of
0.1%, which is equal to the noise level. The DAC resolution is usually chosen
equal to the ADC resolution, or slightly higher, to avoid introducing another
source of quantization error. Once the ADC and DAC resolution have been
selected, the resolution of the reference signal representation must be the same as
that of the ADC and DAC. In fact, if the precision of the reference signal is higher than the ADC resolution, the control error will never go to zero, and therefore
a limit cycle (periodic oscillations associated with nonlinearities) will occur.
Another important design issue, especially for MIMO control, is the choice of
data acquisition system. Ideally, we would use an analog-to-digital converter for
each channel to ensure simultaneous sampling as shown in Figure 12.2a.
However, this approach can be prohibitively expensive, especially for a large
number of channels. A more economical approach is to use a multiplexer
(MUX), with each channel sampled in sequence and the sampled value sent to a
master ADC (Figure 12.2b). If we assume that the sampled signals change relatively slowly, small changes in the sampling instant do not result in significant
errors, and the measured variables appear to be sampled simultaneously. If this
assumption is not valid, a more costly simultaneous sample-and-hold (SSH)
system can be employed, as depicted in Figure 12.2c. The system samples
the channels simultaneously and then delivers the sampled data to the ADC
through a multiplexer. In recent years, the cost of analog-to-digital converters has
decreased significantly, and the use of a simultaneous sample-and-hold system
has become less popular as using an ADC for each channel has become more
affordable.
12.2 Choice of the sampling period
Channel 0
ADC
Controller
Channel n
ADC
(a)
Channel 0
MUX
ADC
Controller
Channel n
(b)
Channel 0
Channel n
SSH
MUX
ADC
Controller
SSH
(c)
FIGURE 12.2
Choices for the data acquisition system. (a) Separate ADC for each channel.
(b) Multiplexer with single master ADC. (c) Simultaneous sample-and-hold, multiplexer,
and single master ADC.
We discuss other considerations related to the choice of the ADC and DAC
components with respect to the sampling period in Section 12.2.2.
12.2 Choice of the sampling period
In Section 2.9, we showed that the choice of the sampling frequency must satisfy
the sampling theorem and is based on the effective bandwidth ωm of the signals
in the control systems. This leads to relation (2.66), where we choose the sampling frequency in the range between 5 and 10 times the value of ωm. We now
discuss the choice of sampling frequency more thoroughly, including the effects
of antialiasing filters, as well as the effects of quantization, rounding, and truncation errors.
12.2.1 Antialiasing filters
If the sampling frequency does not satisfy the sampling theorem (i.e., the sampled
signal has frequency components greater than half the sampling frequency), then
the sampling process creates new frequency components (see Figure 2.9). This
495
496
CHAPTER 12 Practical Issues
External
Input
Computer or
Microprocessor
Analog
System
DAC
ADC
Analog
Output
Antialiasing
Filter
FIGURE 12.3
Control scheme with an antialiasing filter.
phenomenon is called aliasing and must obviously be avoided in a digital control
system. Hence, the continuous signal to be sampled must not include significant
frequency components greater than the Nyquist frequency ωs/2.
For this purpose, it is recommended to low-pass filter the continuous signal
before sampling, especially in the presence of high-frequency noise. The analog
low-pass filter used for this purpose is known as the antialiasing filter. The antialiasing filter is typically a simple first-order RC filter, but some applications
require a higher-order filter such as a Butterworth or a Bessel filter. The overall
control scheme is shown in Figure 12.3.
Because a low-pass filter can slow down the system by attenuating high-frequency dynamics, the cut-off frequency of the low-pass filter must be higher than
the bandwidth of the closed-loop system so as not to degrade the transient
response. A rule of thumb is to choose the filter bandwidth equal to a constant
times the bandwidth of the closed-loop system. The value of the constant varies
depending on economic and practical considerations. For a conservative but more
expensive design, the cut-off frequency of the low-pass filter can be chosen as 10
times the bandwidth of the closed-loop system to minimize its effect on the control system dynamics, and then the sampling frequency can be chosen 10 times
higher than the filter cut-off frequency so there is a sufficient attenuation above
the Nyquist frequency. Thus, the sampling frequency is 100 times the bandwidth
of the closed-loop system. To reduce the sampling frequency, and the associated
hardware costs, it is possible to reduce the antialiasing filter cut-off frequency. In
the extreme case, we select the cut-off frequency slightly higher than the closedloop bandwidth. For a low-pass filter with a high roll-off (i.e., a high-order filter),
the sampling frequency is chosen as five times the closed-loop bandwidth. In
summary, the sampling period T can be chosen (as described in Section 2.9) in
general as
5ωb #
2π
# 100ωb
T
where ωb is the bandwidth of the closed-loop system.
(12.1)
12.2 Choice of the sampling period
If the phase delay introduced by the antialiasing filter is significant, then
(12.1) may not yield good results, and the filter dynamics must be considered
when selecting a dynamic model for the design phase.
EXAMPLE 12.2
Consider a 1 Hz sinusoidal signal of unity amplitude with an additive 50 Hz sinusoidal
noise. Verify the effectiveness of an antialiasing filter if the signal is sampled at a frequency of 30 Hz.
Solution
The noisy 1 Hz analog signal to be sampled is shown in Figure 12.4a. If this signal is sampled
at 30 Hz without an antialiasing filter, the result is shown in Figure 12.4b. Figure 12.4c
shows the filtered analog signal with a first-order antialiasing filter with cut-off frequency
x(t)
x(t)
1
1
0.5
0.5
0
0
–0.5
–0.5
–1
–1
0
0.5
1
1.5
2 2.5
(a)
3
3.5
4t
0
x(t)
1
1
0.5
0.5
0
0
–0.5
–0.5
–1
–1
0.5
1
1.5
2 2.5
(c)
1
1.5
2
2.5
3
3.5
4 t
2.5
3
3.5
4 t
(b)
x(t)
0
0.5
3
3.5
4 t
0
0.5
1
1.5
2
(d)
FIGURE 12.4
Effect of an antialiasing filter on the analog and sampled signals of Example 12.2.
(a) Noisy analog signal. (b) Signal sampled at 30 Hz with no antialiasing filter. (c)
Filtered analog signal with a first-order antialiasing filter with cut-off frequency equal
to 10 Hz. (d) Sampled signal with a first-order antialiasing filter with cut-off
frequency equal to 10 Hz.
497
498
CHAPTER 12 Practical Issues
equal to 10 Hz, and the resulting sampled signal is shown in Figure 12.4d. The sampled sinusoidal signal is no longer distorted because of the use of the antialiasing filter; however, a
small phase delay emerges because of the filter dynamics.
12.2.2 Effects of quantization errors
As discussed in Section 12.1, the design of the overall digital control system
includes the choice of the ADC and DAC components. In this context, the effects
of the quantization due to ADC rounding or truncation (Figure 12.5) are considered in the selection of the sampling period. The noise due to quantization can
be modeled as a uniformly distributed random process with the following mean
and variance values (denoted respectively by e and σ2e ) in the two cases:
Rounding: e 5 0
Truncation: e 5
σ2e 5
q
2
q2
12
σ2e 5
(12.2)
q2
12
(12.3)
where q is the quantization level—namely, the range of the ADC divided by 2n,
and n is the number of bits.
Obviously, the effects of the quantization error increase as q increases and the
resolution of the ADC decreases. To evaluate the influence of q on quantization
noise and on the sampling period, we consider a proportional feedback digital
controller with a gain K applied to the analog first-order lag:
GðsÞ 5
1
τs 1 1
The z-transfer function of the DAC (zero-order hold), analog subsystem, and
ADC (ideal sampler) cascade has the discrete time statespace model
xðk 1 1Þ 5 ½e2T=τ x ðkÞ 1 ½1 2 e2T=τ uðkÞ
yðkÞ 5 xðkÞ
y
y
x
x
q
(a)
q
(b)
FIGURE 12.5
Quantization characteristics of the ADC. (a) Truncating ADC. (b) Rounding ADC.
12.2 Choice of the sampling period
For a truncating ADC as the only source of noise with zero set-point value, the
control action is
uðkÞ 5 2K½yðkÞ 2 wðkÞ
where w is the quantization noise governed by (12.3) and is subtracted from y(k)
with no loss of generality. The statespace model of the closed-loop system is
xðk 1 1Þ 5 e2T=τ 2 Kð1 2 e2T=τ Þ xðkÞ 1 K½1 2 e2T=τ wðkÞ
yðkÞ 5 xðkÞ
For zero initial conditions, the solution of the difference equation is
xðkÞ 5
k21 h
X
e2T=τ 2Kð12e2T=τ Þ
ik2i21
K½1 2 e2T=τ wðkÞ
i50
yðkÞ 5 xðkÞ
The mean value of the output noise is
k21
X
½e2T=τ 2Kð12e2T=τ Þk2i21 K½1 2 e2T=τ EfwðkÞg
i50 0 1
k21
q X
5 K 1 2 e2T=τ @ A ½e2T=τ 2Kð12e2T=τ Þk2i21
2 i50
my ðkÞ 5 EfxðkÞg 5
where E{.} denotes the expectation. If T/τ ,, 1, we use the linear approximation
of the exponential terms e2T/τ 1 2 T/τ to obtain
X
k2i21
T q k21
T
my ðkÞ 5 K
ð11KÞ
12
τ 2 i50
τ
We recall the relationship
N
X
1
ak ; jaj , 1
5
12a
k50
and take the limit as k-N to obtain the steady-state mean
T q
1
K q
ðkÞ
5
K
5
my
T
τ 2 τ ð1 1 KÞ 1 1 K 2
For small gain values, we have
0 1
0 1
K @qA
q
K @ A;
my 5
11K 2
2
For large gain values, we have my q2.
K ,, 1
499
500
CHAPTER 12 Practical Issues
We observe that the mean value is independent of the sampling period, is linear
in the controller gain for small gains, and is almost independent of the gain for large
gains. In any case, the worst mean value is half the quantization interval.
The variance of the output is
σ2y 5 Efx2 ðkÞg 2 E2 fxðkÞg
Using the expression for the mean and after some tedious algebraic manipulations, we can show that
2
K 2 ð12e2T=τ Þ2
q
σ2y 5 5
2T=τ
22T=τ
ðK 1 1Þ½ð1 2 KÞ2Ke
2ðK 1 1Þe
12
If T/τ ,, 1, we use the linear approximations of the exponential terms
e2T/τ 1 2 T/τ and e22T/τ 1 2 2T/τ, and the output variance simplifies to
0 1
8
>
>
KT
q2 A
>
@
>
;
<
K 2 T q2
2τ 12
σ2y 5
>
K 1 1 2τ 12
>
2 τ
>
>
: K 2τ ;
K .. 1
K ,, 1
Unlike the output mean, the output variance is linear in the sampling period
and linear in the controller gain for large gains. Thus, the effect of the quantization noise can be reduced by decreasing the sampling period, once the ADC has
been selected. We conclude that decreasing the sampling period has beneficial
effects with respect to both aliasing and quantization noise. However, decreasing
the sampling period requires more expensive hardware and may aggravate problems caused by the finite-word representation of parameters.
We illustrate this fact with a simple example. Consider an analog controller
with poles at s1 5 21 and s2 5 210. For a digital implementation of the analog
controller with sampling period T 5 0.001, we have the digital controller poles as
given by (6.3) as
z1 5 e20:001 D 0:9990
and
z2 5 e20:01 D 0:9900
If we truncate the two values after two significant digits, we have
z1 5 z2 5 0.99; that is, the two poles of the digital controller become identical and
correspond to two identical poles of the analog controller at
s1 5 s2 5
1
1
ln z1 5 ln z2 D 210:05
T
T
For the longer sampling period T 5 0.1, truncating after two significant digits
gives the poles z1 5 0.90 and z2 5 0.36, which correspond to s121.05 and
s2 5 210.21. This shows that a much better approximation is obtained with the
12.2 Choice of the sampling period
longer sampling period. Example 12.3 shows the effect of a truncating ADC on
the system response.
EXAMPLE 12.3
Consider the process
GðsÞ 5
1
s11
with a proportional digital feedback controller, a sampling period T 5 0.02, and a gain
K 5 5.
(a) Determine the resolution of an 8-bit ADC with a range of 20 V and the resolution of a
14-bit ADC with the same range.
(b) Obtain a plot of the output response and the control input of the closed-loop system
with and without a truncating ADC, first with the 8-bit ADC and then with the 14-bit
ADC, and compare the effect of truncation and ADC resolution on the response.
Solution
(a) The 8-bit ADC has a resolution of 20,000/(28) 5 78.125 mV, while the 14-bit ADC
has a resolution of 20,000/(214) 5 1.22 mV.
(b) Using SIMULINK, we can simulate the closed-loop system to obtain plots of its output
response and control input with and without the quantization effect. For the 8-bit
ADC, the process output is shown in Figure 12.6, and the control input is shown in
Figure 12.7. The differences between the two responses and the two control inputs
are evident. In contrast, for the 14-bit ADC, the results are shown in Figures 12.8 and
12.9, where both the two responses and the two control inputs are almost identical.
0.5
0.45
0.4
Process Output
0.35
0.3
0.25
0.2
0.15
0.1
0.05
0
0
0.1
0.2
0.3
0.4
0.5 0.6
Time s
0.7
0.8
0.9
FIGURE 12.6
Process output with (solid line) and without (dashed line) the 8-bit ADC.
1
501
CHAPTER 12 Practical Issues
0
Controller Output
–0.5
–1
–1.5
–2
–2.5
0
0.1
0.2
0.3
0.4
0.5 0.6
Time s
0.7
0.8
0.9
1
FIGURE 12.7
Controller output with (solid line) and without (dashed line) the 8-bit ADC.
0.5
0.45
0.4
0.35
Process Output
502
0.3
0.25
0.2
0.15
0.1
0.05
0
0
0.1
0.2
0.3
0.4
0.5 0.6
Time s
0.7
0.8
0.9
FIGURE 12.8
Process output with (solid line) and without (dashed line) the 8-bit ADC.
1
12.2 Choice of the sampling period
0
Controller Output
–0.5
–1
–1.5
–2
–2.5
0
0.1
0.2
0.3
0.4
0.5 0.6
Time s
0.7
0.8
0.9
1
FIGURE 12.9
Controller output with (solid line) and without (dashed line) the 8-bit ADC.
12.2.3 Phase delay introduced by the ZOH
As shown in Section 3.3, the frequency response of the zero-order hold can be
approximated as
GZOH ðjωÞ 5
1 2 e2jωT
e2jωT=2
jω
This introduces an additional delay in the control loop approximately equal to
half of the sampling period. The additional delay reduces the stability margins of
the control system, and the reduction is worse as the sampling period is increased.
This imposes an upper bound on the value of the sampling period T.
EXAMPLE 12.4
Let ωc be the gain crossover frequency of an analog control system. Determine the maximum value of the sampling period for a digital implementation of the controller that
decreases the phase margin by no more than 5 .
Solution
Because of the presence of the ZOH, the phase margin decreases by ωcT/2, which yields
the constraint
ωc
or equivalently, T # 0.1745 ωc.
T
π
#5
2
180
503
504
CHAPTER 12 Practical Issues
EXAMPLE 12.5
Consider the tank control system described in Example 2.1 with the transfer function
GðsÞ 5
1:2 21:5s
e
20s 1 1
and the PI controller
CðsÞ 5 7
20s 1 1
20s
Let the actuator and the sensor signals be in the range 0 to 5 V with a sensor gain of
0.169 V/cm. Select a suitable sampling period, antialiasing filter, DAC, and ADC for the
system.
Solution
The gain crossover frequency of the analog control system as obtained using MATLAB is
ωc 5 0.42 rad/s, and the phase margin is φm 5 54 . We select a sampling period T 5 0.2 s
and use a second-order Butterworth antialiasing filter with cut-off frequency of 4 rad/s.
The transfer function of the Butterworth filter is
FðsÞ 5
64
s2 1 11:31s 1 64
The antialiasing filter does not change the gain crossover frequency significantly. The
phase margin is reduced to φm 5 49.6 , which is acceptable. At the Nyquist frequency of
π/T 5 10.47 rad/s, the antialiasing filter decreases the magnitude of the noise by more than
40 dB. The phase delay introduced by the zero-order hold is (ωcT/2) 3 180/π 5 3.6 , which is
also acceptable. We select a 12-bit ADC with a quantization level of 1.2 mV, which corresponds
to a quantization error in the fluid level of 0.07 mm. We also select a 12-bit DAC. Because the
conversion time is on the order of microseconds, this does not influence the overall design.
12.3 Controller structure
Section 12.2 demonstrates how numerical errors can affect the performance of a
digital controller. To reduce numerical errors and mitigate their effects, we must
select an appropriate controller structure for implementation. To examine the
effect of controller structure on errors, consider the controller
CðzÞ 5
Nðz; qÞ bT zm
5 T a
a zn
Dðz; qÞ
a 5 ½a0 ðqÞ a1 ðqÞ ? an ðqÞT
b 5 ½b0 ðqÞ b1 ðqÞ ? bm ðqÞT
zl 5 ½1 z ? zl (12.4)
where q is an l 3 1 vector of controller parameters. If the nominal parameter vector is q and the corresponding poles are pi , i 5 1, 2, . . . n, for an nth-order controller, then the nominal characteristic equation of the controller is
Dðpi ; q Þ 5 0; i 5 1; 2; . . .; n
(12.5)
12.3 Controller structure
In practice, the parameter values are only approximately implemented and the
characteristic equation of the system is
3
3T
@D
@D
Dðz; qÞ Dðpi ; q Þ 1 5
δpi 1 5
δa
@z
@a
2
z5pi
3
@D5
5
@z
3T
@D
δpi 1 5
@a
δa 0
z5pi
z5pi
@D 4 @D
5
@a
@a0
z5pi
@D
@a1
?
3T
@D 5
@an
(12.6)
In terms of the controller parameters, the perturbed characteristic equation is
3
3
@D5
@DT @a5
δpi 1
δq 0
Dðz; qÞ @z
@a @q
z5pi
2 3
@a
@ai
54 5
@qj
@q
q5q
(12.7)
We solve for parameter perturbations in the location of the ith pole
@DT @a
δq
δpi 5 2
@a @q q5q @D
@z z5p
(12.8)
i
To characterize the effect of a particular parameter on the ith pole, we can set
the perturbations in all other parameters to zero to obtain
3
T
@D
@a
δqi
5
δpi 5 2
3
@a @qi
@D
q5q
5
@z
z5pi
"
#T
@a0 @a1
@an
@a
5 @q @q ? @q
(12.9)
i
i
i
@qi
This concept is explained by the following example. Consider the following
second-order general controller:
CðzÞ 5
ða 1 bÞz 2ðap2 1 bp1 Þ
z2 2ðp1 1 p2 Þz 1 p1 p2
505
506
CHAPTER 12 Practical Issues
Let a1 5 2(p1 1 p2) and a0 5 p1p2 denote the nominal coefficients of the characteristic equation of the controller, and write the characteristic equation as
Dðz; a1 ; a0 Þ 5 0
When a coefficient λi is changed (due to numerical errors) to λi 1 δλi, then
the position of a pole is changed according to the following equation (where second- and higher-order terms are neglected):
@D @D
δpi 1
δλi
Dðpi 1 δpi ; λi 1 δλi Þ 5 Dðpi ; λi Þ 1
@z z5pi
@λi
That is,
@D
δpi
@λ
i
52
@D δλi
@z z5pi
Now we have the partial derivatives
@D
5 2z 1 a1
@z
@D
5z
@a1
@D
51
@a0
and therefore
δp1
p1
p1
5
52
;
δa1
2p1 2ðp1 1 p2 Þ p1 2 p2
δp2
p2
p2
5
52
δa1
2p2 2ðp1 1 p2 Þ p2 2 p1
δp1
1
1
5
52
;
δa0
2p1 2ðp1 1 p2 Þ p1 2 p2
δp2
1
1
5
52
δa0
2p2 2ðp1 1 p2 Þ p2 2 p1
Thus, the controller is most sensitive to changes in the last coefficient of the
characteristic equation, and its sensitivity increases when the poles are close. This
concept can be generalized to high-order controllers. Note that decreasing the
sampling period draws the poles closer when we start from an analog design. In
fact, for T-0 we have that
pz 5 eps T -1
independently on the value of the analog pole ps.
These problems can be avoided by writing the controller in an equivalent
parallel form:
a
b
1
CðzÞ 5
z 2 p1
z 2 p2
We can now analyze the sensitivity of the two terms of the controller separately to show that the sensitivity is equal to one, which is less than the previous
case if the poles are close. Thus, the parallel form is preferred. Similarly, the parallel form is also found to be superior to the cascade form:
CðzÞ 5
ða 1 bÞz 2ðap2 1 bp1 Þ
1
3
z 2 p1
z 2 p2
12.4
PID control
EXAMPLE 12.6
Write the difference equations in direct, parallel, and cascade forms for the system
CðzÞ 5
UðzÞ
z 2 0:4
5 2
EðzÞ
z 2 0:3z 1 0:02
Solution
The difference equation corresponding to the direct form of the controller is
uðkÞ 5 0:3uðk 2 1Þ 2 0:02uðk 2 2Þ 1 eðk 2 1Þ 2 0:4eðk 2 2Þ
For the parallel form, we obtain the partial fraction expansion of the transfer function
C ðzÞ 5
U ðzÞ
22
3
5
1
E ðzÞ
z 2 0:2 z 2 0:1
This is implemented using the following difference equations:
u1 ðkÞ 5 0:2uðk 2 1Þ 2 2eðk 2 1Þ
u2 ðkÞ 5 0:1uðk 2 1Þ 1 3eðk 2 1Þ
uðkÞ 5 u1 ðkÞ 1 u2 ðkÞ
Finally, for the cascade form we have
CðzÞ 5
UðzÞ z 2 0:4 1
XðzÞ UðzÞ
5
5
EðzÞ
z 2 0:2 z 2 0:1 EðzÞ XðzÞ
which is implemented using the difference equations
xðkÞ 5 0:2xðk 2 1Þ 1 eðkÞ 2 0:4eðk 2 1Þ
uðkÞ 5 0:1uðk 2 1Þ 1 xðk 2 1Þ
12.4 PID control
In this section, we discuss several critical issues related to the implementation of
PID controllers. Rather than providing an exhaustive discussion, we highlight a
few problems and solutions directly related to digital implementation.
12.4.1 Filtering the derivative action
The main problem with derivative action is that it amplifies the high-frequency
noise and may lead to a noisy control signal that can eventually cause serious
damage to the actuator. It is therefore recommended that one filter the overall
control action with a low-pass filter or, alternatively, filter the derivative action.
In this case, the controller transfer function in the analog case can be written as
(see (5.20))
!
1
Td s
1
CðsÞ 5 Kp 1 1
Ti s 1 1 TNd s
507
508
CHAPTER 12 Practical Issues
where N is a constant in the interval [1, 33], Kp is the proportional gain, Ti is the
integral time constant, and Td is the derivative time constant.
In the majority of cases encountered in practice, the value of N is in the smaller interval [8, 16]. The controller transfer function can be discretized as discussed
in Chapter 6. A useful approach in practice is to use the forward differencing
approximation for the integral part and the backward differencing approximation
for the derivative part. This gives the discretized controller transfer function
0
1
B
CðzÞ 5 Kp B
@1 1
T
1
Ti ðz 2 1Þ
z21 C
C
U
Td
Td A
T1
z2
N
NT 1 Td
Td
(12.10)
which can be simplified to
CðzÞ 5
where
K0 1 K1 z 1 K2 z2
ðz 2 1Þðz 2 γÞ
1
T
T
T
NT
d
d
d A
K0 5 Kp @
2
1
NT 1 Td
NT 1 Td
Ti NT 1 Td
0
1
T
T
NT
d
d
A
2 12
K 1 5 2 K p @1 1
NT 1 Td
NT 1 Td
Ti
0
1
NT
d A
K2 5 Kp @1 1
NT 1 Td
(12.11)
0
γ5
(12.12)
Td
1
5
NT 1 Td
NðT=Td Þ 1 1
EXAMPLE 12.7
Select a suitable derivative filter parameter value N for the PID controller described in
Example 5.9 with a sampling period T 5 0.01, and obtain the corresponding discretized
transfer function of (12.11).
Solution
The analog PID parameters are Kp 5 2.32, Ti 5 3.1, and Td 5 0.775. We select the filter
parameter N 5 20 and use (12.12) to obtain K0 5 38.72, K1 5 277.92, K2 5 39.20, and
γ 5 0.79.
The discretized PID controller expression is therefore
CðzÞ 5
K0 1 K1 z 1 K2 z2
38:72 2 77:92z 1 39:20z2
5
ðz 2 1Þðz 2 γÞ
ðz 2 1Þðz 2 0:79Þ
12.4
PID control
12.4.2 Integrator windup
Most control systems are based on linear models and design methodologies.
However, every actuator has a saturation nonlinearity, as in the control loop
shown in Figure 12.10, which affects both the analog and digital control. The
designer must consider the nonlinearity at the design stage to avoid performance
degradation. A common phenomenon related to the presence of actuator saturation is known as integrator windup. If not properly handled, it may result in a
step response with a large overshoot and settling time.
In fact, if the control variable attains its saturation limit when a step input is
applied, the control variable becomes independent of the feedback and the system
behaves as in the open-loop case. The control error decreases more slowly than in
the absence of saturation, and the integral term becomes large or winds up. The
large integral term causes saturation of the control variable even after the process
output attains its reference value and a large overshoot occurs.
Many solutions have been devised to compensate for integrator windup and
retain linear behavior. The rationale of these anti-windup techniques is to design
the control law, disregarding the actuator nonlinearity, and then compensate for
the detrimental effects of integrator windup.
One of the main anti-windup techniques is the so-called conditional integration, which keeps the integral control term constant when a specified condition is
met. For example, the integral control term is kept constant if the integral component of the computed control exceeds a given threshold specified by the designer.
Alternatively, the integral controller is kept constant if the actuator saturates with
the control variable and the control error having the same sign (i.e., if u e . 0).
The condition u e . 0 implies that the control increases rather than corrects the
error due to windup and that integral action should not increase.
On the other hand, a positive saturation with u e , 0 means that the error is
negative and therefore the integral action is decreasing and there is no point in
keeping it constant. The same reasoning can be easily applied in case of a negative saturation. Thus, the condition avoids inhibiting the integration when it helps
to push the control variable away from saturation.
An alternative technique is back-calculation, which reduces (increases) the
integral control when the maximum (minimum) saturation limit is attained by
adding to the integrator a term proportional to the difference between the
Reference
Input
Controller
FIGURE 12.10
Control loop with actuator saturation.
v
u
Process
Output y
Process
509
510
CHAPTER 12 Practical Issues
computed value of the control signal v and its saturated value u. In other words,
the integral value I(k) is determined by
IðkÞ 5 Iðk 2 1Þ 1
Kp
1
eðkÞ 2 ðvðkÞ 2 uðkÞÞ
Ti
Tt
(12.13)
where Tt is the tracking time constant. This is a tuning parameter that determines the rate at which the integral term is reset.
EXAMPLE 12.8
Consider the digital proportional-integral (PI) controller transfer function
CðzÞ 5
1:2z 2 1:185
z21
with Kp 5 1.2, Ti 5 8, sampling period T 5 0.1, and the process transfer function
GðsÞ 5
1
e25s
10s 1 1
Obtain the step response of the system (1) if the saturation limits of the actuator are
umin 5 21.2 and umax 5 1.2, and (2) with no actuator saturation. Compare and discuss the
two responses, and then use back-calculation to reduce the effect of windup on the step
response.
Solution
For the system with actuator saturation and no anti-windup strategy, the process output y,
controller output v, and process input u are shown in Figure 12.11. We observe that the
actuator output exceeds the saturation level even when the process output attains its
2
v
1.8
u
1.6
y
1.4
1.2
1
0.8
0.6
0.4
0.2
0
0
10
20
30
40
50 60
Time s
70
80
90 100
FIGURE 12.11
Process input u, controller output v, and process output y with actuator saturation
and no anti-windup strategy.
12.4
2
PID control
v
1.8
1.6
y
1.4
1.2
1
0.8
0.6
0.4
0.2
0
0
10
20
30
40
50 60
Time s
70
80
90 100
FIGURE 12.12
Controller output v and process output y with no actuator saturation.
1.4
u
1.2
1
v
0.8
y
0.6
0.4
0.2
0
0
10
20
30
40
50 60
Time s
70
80
90 100
FIGURE 12.13
Process input u, controller output v, and process output y with actuator saturation
and a back-calculation anti-windup strategy.
reference value, which leads to a large overshoot and settling time. The response after the
removal of saturation nonlinearity is shown in Figure 12.12. The absence of saturation
results in a faster response with fast settling to the desired steady-state level.
The results obtained by applying back-calculation with Tt 5 Ti 5 8 are shown in
Figure 12.13. The control variable is kept at a much lower level, and this helps avoid the
overshoot almost entirely. The response is slower than the response with no saturation but
is significantly faster than the response with no anti-windup strategy.
511
512
CHAPTER 12 Practical Issues
12.4.3 Bumpless transfer between manual and automatic mode
When the controller can operate in either manual mode or automatic mode,
switching between the two modes of operation must be handled carefully to avoid
a bump in the process output at the switching instant. During manual mode, the
operator provides feedback control and the automatic feedback is disconnected.
The integral term in the feedback controller can assume a value different from the
one selected by the operator. Simply switching from automatic to manual, or vice
versa, as in Figure 12.14, leads to a bump in the control signal, even if the control
error is zero. This results in an undesirable bump in the output of the system.
For a smooth or bumpless transfer between manual control and the automatic
digital controller C(z), we use the digital scheme shown in Figure 12.15. We write
the automatic controller transfer function in terms of an asymptotically
stable controller D(z) with unity DC gain—that is, D(1) 5 1—as
CðzÞ 5
K
1 2 DðzÞ
(12.14)
We then solve for D(z) in terms of C(z) to obtain
DðzÞ 5
R(z)
+
M
E(z)
C(z)
−
CðzÞ 2 K
CðzÞ
Y(z)
A U(z)
DAC
G(s)
ADC
FIGURE 12.14
Block diagram for bumpy manual (M)/automatic (A) transfer.
C(z)
D(z)
M
e(k)
A
K
+
+
FIGURE 12.15
Block diagram for bumpless manual (M)/automatic (A) transfer.
M
A
u(k)
12.4
PID control
If C(z) is the PID controller transfer function of (12.11), we have
DðzÞ 5
ðK2 2 KÞz2 1ðK1 1 K 1 KγÞz 1 K0 2 Kγ
K2 z2 1 K1 z 1 K0
(12.15)
If the coefficient of the term z2 in the numerator is nonzero, then the controller
has the form
DðzÞ 5
UðzÞ
5 ðK2 2 KÞ 1 Da ðzÞ
EðzÞ
where Da(z) has a first-order numerator polynomial. The controller output
u(k 1 2) is equal to the sum of two terms, one of which is the controller output
u(k 1 2) itself. Thus, the solution of the related difference equation cannot be
computed by a simple recursion. This undesirable controller structure is known as
an algebraic loop. To avoid an algebraic loop, we impose the condition
K 5 K2
(12.16)
2
to eliminate the z term in the numerator. We illustrate the effectiveness of the
bumpless manual/automatic mode scheme using Example 12.9.
EXAMPLE 12.9
Verify that a bump occurs if switching between manual and automatic operation uses the
configuration shown in Figure 12.14 for the process
GðsÞ 5
1
e22s
10s 1 1
and the PID controller (T 5 0.1)
CðzÞ 5
44z2 2 85:37z 1 41:43
z2 2 1:368z 1 0:368
Design a scheme that provides a bumpless transfer between manual and automatic
modes.
Solution
The unit step response for the system shown in Figure 12.14 is shown in Figure 12.16,
together with the automatic controller output. The transfer between manual mode, where a
step input u 5 1 is selected, and automatic mode occurs at time t 5 100. The output of the
PID controller is about 11, which is far from the reference value of unity at t 5 100. This
leads to a significant bump in the control variable on switching from manual to automatic
control that acts as a disturbance, causing a bump in the process output. To eliminate the
bump, we use the bumpless transfer configuration shown in Figure 12.15 with the PID
controller parameters
K2 5 44; K1 5 285:37; K0 5 41:43; γ 5 0:368
Using (12.15) and (12.16), we have
DðzÞ 5
2 0:572z 1 0:574
z2 2 1:94z 1 0:942
513
Process Output
CHAPTER 12 Practical Issues
3
2
1
0
0
50
100
150
100
150
Controller Output
Time s
(a)
40
20
0
0
50
Time s
(b)
FIGURE 12.16
Process Output
Process output (a) and controller output (b) for the system without bumpless
transfer.
Controller Output
514
1
0.5
0
0
50
0
50
Time s
(a)
100
150
100
150
1
0.5
0
Time s
(b)
FIGURE 12.17
Process output (a) and controller output (b) for the bumpless manual/automatic
transfer shown in Figure 12.15.
and K 5 44. The results of Figure 12.17 show a bumpless transfer, with the PID output at
t 5 100 equal to one.
12.4
PID control
12.4.4 Incremental form
Integrator windup and bumpless transfer issues are solved by implementing the
PID controller in incremental form. We determine the increments in the control
signal at each sampling period instead of determining the actual values of the
control signal. This moves the integral action outside the control algorithm. To
better understand this process, we consider the difference equation of a PID controller (the filter on the derivative action is not considered for simplicity):
!
k
TX
Td
eðiÞ 1 ðeðkÞ 2 eðk 2 1ÞÞ
uðkÞ 5 Kp eðkÞ 1
T
Ti i50
Subtracting the expression for u(k 2 1) from that of u(k), we obtain the
increment
T
Td
2Td
Td
eðkÞ 1 21 2
eðk 2 1Þ 1 eðk 2 2Þ
11 1
uðkÞ 2 uðk 2 1Þ 5 Kp
T
T
T
Ti
which can be rewritten more compactly as
uðkÞ 2 uðk 2 1Þ 5 K2 eðkÞ 1 K1 eðk 2 1Þ 1 K0 eðk 2 2Þ
where
T
Td
K2 5 Kp 1 1 1
T
Ti
2Td
K1 5 2 Kp 1 1
T
(12.17)
K0 5 Kp
Td
T
(12.18)
(12.19)
(12.20)
From the difference equation (12.17), we can determine the increments in the
control signal at each sampling period. We observe that in (12.17) there is no
error accumulation (in this case the integral action can be considered “outside”
the controller), and the integrator windup problem does not occur. In practice, it
is sufficient that the control signal is not incremented when the actuator
saturates—namely, we have u(k) 5 u(k 2 1) when v(k) 5 u(k) in Figure 12.10.
Further, the transfer between manual mode and automatic mode is bumpless as
long as the operator provides the increments in the control variable rather than
their total value.
The z-transform of the difference equation (12.17) gives the PID controller
z-transfer function in incremental form as
CðzÞ 5
∆UðzÞ K2 z2 1 K1 z 1 K0
5
z2
EðzÞ
where ∆U(z) is the z-transform of the increment u(k) 2 u(k 2 1).
(12.21)
515
CHAPTER 12 Practical Issues
EXAMPLE 12.10
For the process described in Example 12.8 and an analog PI controller with Kp 5 1.2 and
Ti 5 8, verify that windup is avoided with the digital PI controller in incremental form
(T 5 0.1) if the saturation limits of the actuator are umin 5 21.2 and umax 5 1.2.
Solution
Using (12.18) and (12.19), we obtain K2 5 1.215 and K1 5 21.2 and the digital controller
transfer function
CðzÞ 5
1:215z 2 1:2
z21
The output obtained with the PI controller in incremental form (avoiding controller
updating when the actuator saturates) is shown in Figure 12.18. Comparing the results to
those shown in Figure 12.11, we note that the response is much faster both in rising to
the reference value and in settling. In Figure 12.18, the effects of saturation are no longer
noticeable.
1
0.8
Process Output
516
0.6
0.4
0.2
0
0
10
20
30
40
50 60
Time s
70
80
90 100
FIGURE 12.18
Process output with the PID in incremental form.
12.5 Sampling period switching
In many control applications, it is necessary to change the sampling period during
operation to achieve the optimal usage of the computational resources. In fact, a
single CPU usually performs many activities such as data storage, user interface,
and communication, and possibly implements more than one controller. It is
therefore necessary to optimize CPU utilization by changing the sampling period.
For a given digital control law, on the one hand, it is desirable to decrease the
sampling period to avoid performance degradation, but on the other hand,
decreasing it can overload the CPU and violate real-time constraints.
12.5 Sampling period switching
The problem of changing the control law when the sampling frequency changes
can be solved by switching between controllers working in parallel, each with a
different sampling period. This is a simple task provided that bumpless switching
is implemented (see Section 12.4.3). However, using multiple controllers in parallel is computationally inefficient and is unacceptable if the purpose is to optimize
CPU utilization. Thus, it is necessary to shift from one controller to another when
the sampling period changes rather than operate controllers in parallel.
This requires computing the new controller parameters as well as past values
of error and control variables that it needs to compute the control before switching. If the original sampling interval is T0 and the new sampling interval is T,
then we must switch from the controller
C 0 ðzÞ 5
UðzÞ bT zm
5 T
a zn
EðzÞ
a 5 a0 ðT 0 Þ
b 5 b0 ðT 0 Þ
zl 5 1
a1 ðT 0 Þ ? 1
T
b1 ðT 0 Þ ? bm ðT 0 Þ
z ?
(12.22)
T
zl to the controller
CðzÞ 5
UðzÞ bT zm
5 T
a zn
EðzÞ
a 5 a0 ðTÞ
a1 ðTÞ
? 1 T
b 5 b0 ðTÞ b1 ðTÞ ?
zl 5 1 z
bm ðTÞ T
? zl Equivalently, we switch from the difference equation
uðkT 0 Þ 5 2an21 ðT 0 Þuððk 2 1ÞT 0 Þ 2 . . . 2 a0 ðT 0 Þuððk 2 nÞT 0 Þ
1 b0 ðT 0 Þeððk 2 n 1 mÞT 0 Þ 1 . . . 1 bm ðT 0 Þeððk 2 nÞT 0 Þ
to the difference equation
uðkTÞ 5 2an21 ðTÞuððk 2 1ÞTÞ 2 . . . 2 a0 ðTÞuððk 2 nÞTÞ
1 b0 ðTÞeððk 2 n 1 mÞTÞ 1 . . . 1 bm ðTÞeððk 2 nÞTÞ
Thus, at the switching time instant, we must recompute the values of the parameter vectors a and b, as well as the corresponding past m 1 1 values of the tracking
error e and the past n values of the control variable u.
We compute the new controller parameters using the controller transfer function, which explicitly depends on the sampling period. For example, if the PID
controller of (12.11) is used, the new controller parameters can be easily computed using (12.12) with the new value of the sampling period T, or equivalently,
using (12.10) where the sampling period T appears explicitly. To compute the
517
518
CHAPTER 12 Practical Issues
past values of the tracking error e and of the control variable u, different techniques can be applied depending on whether the sampling period increases or
decreases. For simplicity, we only consider the case where one sampling period is
a divisor or multiple of the other. The case where the ratio between the previous
and the new sampling periods (or vice versa) is not an integer is a simple extension, which is not considered here.
If the new sampling period T is a fraction of the previous sampling period T0
(i.e., T0 5 λT), the previous n values of the control variable [u((k 2 1)T), u((k 2 2)T),
. . ., u((k 2 n)T)] are determined with the control variable kept constant during the
past λ periods. The m 1 1 previous error values are computed using an interpolator
such as a cubic polynomial. In particular, the coefficients c3, c2, c1, and c0 of a thirdorder polynomial e~ ðtÞ 5 c3 t3 1 c2 t2 1 c1 t 1 c0 can be determined by considering
the past three samples and the current value of the control error. The data yield the
following linear system:
2
32 3 2
3
c3
eððk 2 3ÞT 0 Þ
ððk23ÞT 0 Þ3 ððk23ÞT 0 Þ2 ðk 2 3ÞT 0 1
6 ððk22ÞT 0 Þ3 ððk22ÞT 0 Þ2 ðk 2 2ÞT 0 1 76 c2 7 6 eððk 2 2ÞT 0 Þ 7
6
76 7 6
7
4 ððk21ÞT 0 Þ3 ððk21ÞT 0 Þ2 ðk 2 1ÞT 0 1 54 c1 5 5 4 eððk 2 1ÞT 0 Þ 5 (12.23)
c0
eðkT 0 Þ
ðkT 0 Þ2
kT 0
1
ðkT 0 Þ3
Once the coefficients of the polynomial function have been determined, the
previous values of the control error with the new sampling period are the values
of the polynomial functions at the required sampling instants. The procedure is
illustrated in Figure 12.19, where the dashed line connecting the control error
values between (k 2 3)T and kT is the polynomial function e~ ðtÞ.
If the new sampling period T is a multiple of the previous sampling period T 0
(i.e., T 5 λT 0 ), the previous m error samples are known. However, the memory
u
e
e~
t
(k−6)T
(k−2)T ′
(k−5)T
(k−4)T
(k−3)T
(k−1)T ′
(k−2)T
(k−1)T
kT
kT ′
FIGURE 12.19
Switching controller sampling from a sampling period T 0 to a faster sampling rate with
sampling period T 5 T 0 /3.
12.5 Sampling period switching
buffer must be large enough to store them. If the pole-zero difference of the process is equal to one, the equivalent n past control actions are approximately computed as the outputs estimated using the model of the control system with
sampling period T. Specifically, let the process model obtained by discretizing the
analog process with sampling period T be
GZAS ðzÞ 5
YðzÞ
β zh21 1 β h22 zh22 1 . . . 1 β 0
5 h21h
z 1 αh21 zh21 1 . . . 1 α0
UðzÞ
where h $ n. Then the equivalent past n control actions u((k 2 1)T), u((k 2 2)T), . . . ,
u((k 2 n)T) are determined by minimizing the difference between the measured output and that estimated by the model at the switching time:
minjyðkTÞ 2 ð2 αh21 yððk 2 1ÞTÞ 2 . . . 2 α0 yððk 2 hÞTÞ
1 β h21 uððk 2 1ÞTÞ 1 . . .β 0 uððk 2 hÞTÞÞj
(12.24)
We solve the optimization problem numerically using an appropriate approach
such as the simplex algorithm (see Section 12.5.1). To initiate the search, initial
conditions must be provided. The initial conditions can be selected as the values
of the control signal at the same sampling instants determined earlier with the faster controller. The situation is depicted in Figure 12.20.
12.5.1 MATLAB commands
When the sampling frequency is increased, the array en of the m 1 1 previous
error values can be computed using the following MATLAB command:
.. en 5 interp1ðts; es; tn;0 cubic0 Þ
u
e
t
(k−6)T ′
(k−2)T
(k−5)T ′
(k−4)T ′
(k−3)T ′
(k−1)T
(k−2)T ′
(k−1)T ′
kT ′
kT
FIGURE 12.20
Switching controller sampling from a sampling period T0 to a slower sampling rate with
sampling period T 5 3T0 .
519
520
CHAPTER 12 Practical Issues
where ts is a vector containing the last four sampling instants [(k 2 3)T 0 , (k 2 2)
T 0 , (k 2 1)T 0 , kT 0 ] of the slower controller and es is a vector containing the corresponding control errors [e((k 2 3)T 0 ), e((k 2 2)T 0 ), e((k 2 1)T 0 ), e(kT 0 )]. The array
tn contains the m 1 1 sampling instants for which the past control errors for the
new controller must be determined.
Alternatively, the vector of the coefficients c 5 [c3, c2, c1, c0]T of the cubic
polynomial can be obtained by solving the linear system (12.23) using the
command
.. c 5 linsolveðM; eÞ
where M is the matrix containing the time instants and e is the column vector of
error samples such that (12.23) is written as M c 5 e. Once the coefficients of the
polynomial are computed, the values of the control error at previous sampling
instants can be easily determined by interpolation.
To find the past control values using the simplex algorithm, use the command
.. un 5 fminsearchð@ðuÞðabsðyk 2 ah1 yk1 2 . . . 2 a0 ykh
1 bh1 uð1Þ 1 . . . 1 b0 uðnÞÞÞ;initcondÞ
where initcond is the vector of n elements containing the initial conditions, un is
the vector of control variables [u((k 2 1)T), u((k 2 2)T), . . . , u((k 2 n)T)], and the
terms in the abs function are the values corresponding to expression (12.24), with
the exception of u(1), . . . , u(n) that must be written explicitly. Details are provided in the following examples.
EXAMPLE 12.11
Discuss the effect of changing the sampling rate on the step response of the process
described in Example 12.8 and an analog PI controller with Kp 5 1.2 and Ti 5 8. Initially,
use a digital PI controller in incremental form with T0 5 0.4 s; then switch at time t 5 22 s
to a faster controller with T 5 0.1 s.
Solution
For PI control, we set Td to zero in (12.18) to (12.20) to obtain the parameters
0
1
0
T
K2 5 Kp @1 1 A 5 1:2ð1 1 0:4=8Þ 5 1:26
Ti
K1 5 2Kp 5 21:2
K0 5 0
Using (12.21), the initial digital controller transfer function is
CðzÞ 5
UðzÞ 1:26z 2 1:2
5
EðzÞ
z21
We simulate the system using MATLAB to compute the control and error signals. The
control signal at t 5 20.8, 21.2, and 21.6 s is plotted in Figure 12.21, whereas the error at
time t 5 20.8, 21.2, 21.6, and 22 s is plotted as circles in Figure 12.22. At time t 5 22 s,
12.5 Sampling period switching
0.88
0.87
Control Signal
0.86
0.85
0.84
0.83
0.82
0.81
0.8
20.8
21
21.2
21.4
Time s
21.6
21.8
22
FIGURE 12.21
Control signal (solid line) obtained by simulating the system of Example 12.11 with
sampling period T 0 5 0.4 s. Circles denote the control values at the sampling instants and
stars denote the control values for T 5 0.1 s.
–0.3125
–0.313
Error
–0.3135
–0.314
–0.3145
–0.315
–0.3155
–0.316
20.8
21
21.2
21.4
Time s
21.6
21.8
22
FIGURE 12.22
Control error interpolation eðtÞ
~ (solid line) obtained by simulating the system of
Example 12.10 with sampling period T 0 5 0.4 s. Circles denote the error values at the
sampling instants and stars denote the error values for T 5 0.1 s.
521
522
CHAPTER 12 Practical Issues
the sampling period switches to T 5 0.1 s and the control signal must be recomputed. The
new PI controller parameters are
0
1
T 0A
@
K2 5 Kp 1 1
5 1:2ð1 1 0:1=8Þ 5 1:215
Ti
K1 5 2 Kp 5 2 1:2
K0 5 0
yielding the following transfer function:
CðzÞ 5
1:215z 2 1:2
z21
The associated difference equation is therefore
uðkÞ 5 uðk 2 1Þ 1 1:215eðkÞ 2 1:2eðk 2 1Þ
(12.25)
Whereas at t 5 22 s the value of e(k) 5 20.3139 is unaffected by switching, the values
of u(k 2 1) and e(k 2 1) must in general be recomputed for the new sampling period. For a
first-order controller, the value of u(k 2 1) is the value of the control signal at time
t 5 22 2 T 5 22 2 0.1 5 21.9 s. This value is the same as that of the control signal for sampling period T0 over the interval 21.6 to 22 s. From the simulation results shown in
Figure 12.22, we obtain u(k 2 1) 5 0.8219, where the values of the control signal at
t 5 21.7, 21.8, and 21.9 s with the new controller are denoted by stars.
To calculate the previous value of the control error e(k 2 1), we use a cubic polynomial
to interpolate the last four error values with the slower controller (the interpolating function
is the solid line in Figure 12.22). Using the control errors at t 5 21.7, 21.8, 21.9 s,
e(21.6) 5 20.3155, e(21.2) 5 20.3153, and e(20.8) 5 20.3131, the linear system (12.23) is
solved for the coefficients c3 5 20.0003, c2 5 0.0257, c1 5 20.6784, and c0 5 5.4458. To
solve the linear system, we use the MATLAB commands
.. M 5 [20.8^3 20.8^2 20.8 1; 21.2^3 21.2^2 21.2 1; 21.6^3 21.6^2
21.6 1; 22.0^3 22.0^2 22.0 1];
.. e 5 [ 2 0.3131 2 0.3153 2 0.3155 2 0.3139]’;
.. c 5 linsolve(M,e); % Solve the linear system M c 5 e
.. en 5 c(1) 21.9^3 1 c(2) 21.9^2 1 c(3) 21.9 1 c(4);
The value of the control error at time t 5 21.9 s is therefore
eðk 2 1Þ 5 2 0:0003U21:93 1 0:0257U21:92 2 0:6784U21:9 1 5:4458 5 2 0:3144
Alternatively, we can use the MATLAB command for cubic interpolation
.. en 5 interp1ð½208 21:2 21:6 22;½ 2 0:3131 2 0:3153 2 0:3155 2 0:3139; 21:9;0 cubic0 Þ;
The control error for the faster controller at t 5 21.7, 21.8, and 21.9 s is denoted by
stars in Figure 12.22. From (12.25), we observe that only the error at t 5 21.9 is needed to
calculate the control at t 5 22 s after switching to the faster controller. We compute the
control value:
uðkÞ 5 uðk 2 1Þ 1 1:215eðkÞ 2 1:2eðk 2 1Þ
5 0:8219 1 1:215 3 ð2 0:3139Þ 2 1:2 3 ð2 0:3144Þ 5 0:8178
We compute the control at time t 5 22.1 s and the subsequent sampling instants using
the same expression without the need for further interpolation. Note that the overall performance is not significantly affected by changing the sampling period, and the resulting
process output is virtually the same as the one shown in Figure 12.12.
12.5 Sampling period switching
EXAMPLE 12.12
Design a digital controller for the DC motor speed control system described in Example
6.17 with transfer function
GðsÞ 5
1
ðs 1 1Þðs 1 10Þ
to implement the analog PI controller
CðsÞ 5 47:2
s11
s
with the sampling period switched from T 0 5 0.01 s to T 5 0.04 s at time t 5 0.5 s. Obtain
the step response of the closed-loop system, and discuss your results.
Solution
Applying the bilinear transformation with T 0 5 0.01 to the controller transfer function C(s),
we obtain the initial controller transfer function
CðzÞ 5
47:44z 2 46:96
z21
We simulate the system using MATLAB to compute the error and process output
values. The error values at t 5 0.42, . . . , 0.5 s are shown in Figure 12.23. The process output for a unit step reference input step at the same sampling instants is shown in
Figure 12.24.
Starting at t 5 0.5 s, the controller transfer function obtained by bilinearly transforming
the analog controller with sampling period T 5 0.04 s becomes
CðzÞ 5
UðzÞ 48:14z 2 46:26
5
EðzÞ
z21
0.07
0.06
0.05
Error
0.04
0.03
0.02
0.01
0
–0.01
–0.02
0.42 0.43 0.44 0.45
0.46 0.47 0.48 0.49
Time s
0.5
FIGURE 12.23
Control error for the two controllers described in Example 12.12. Circles denote the error
values for the faster controller; the dark circles denote the error values for the slower
controller.
523
CHAPTER 12 Practical Issues
1.03
1.02
1.01
Process Output
524
1
0.99
0.98
0.97
0.96
0.95
0.42
0.43
0.44
0.45
0.46 0.47
Time s
0.48
0.49
0.5
FIGURE 12.24
Process output for the faster controller described in Example 12.12.
The corresponding difference equation
uðkÞ 5 uðk 2 1Þ 1 48:14eðkÞ 2 46:26eðk 2 1Þ
is used to calculate the control variable starting at t 5 0.5 s. From the MATLAB simulation
results, the error values needed to compute the control at t 5 0.5 s are e(k) 5 20.0172 at
t 5 0.5 s and e(k 2 1) 5 0.0151 at t 5 0.5 2 0.04 5 0.46 s. The control u(k 2 1) at t 5 0.46 s
must be determined by solving the optimization problem (12.24). The z-transfer function
of the plant, ADC and DAC, with T 5 0.04 is
GZAS ðzÞ 5
5
YðzÞ
6:936z 1 5:991
5 1024 2
UðzÞ
z 2 1:631z 1 0:644
6:936 3 1024 z21 1 5:991 3 1024 z22
1 2 1:631z21 1 0:644z22
and the corresponding difference equation is
yðkÞ 5 1:631yðk 2 1Þ 2 0:644yðk 2 2Þ 1 6:936U1024 uðk 2 1Þ 1 5:991U1024 uðk 2 2Þ
Therefore, the optimization problem is
minjyð0:5Þ 2 1:631yð0:46Þ 2 0:644yð0:42Þ 1 6:936 3 1024 uð0:46Þ 1 5:991 3 1024 uð0:42Þ j
Using the output values y(0.5) 5 1.0235, y(0.46) 5 0.9941, and y(0.42) 5 0.9522, the
values of u(0.46) and u(0.42) are computed by solving the optimization problem using the
following MATLAB command:
.. u 5 fminsearch(@(un)(abs(1.0235 2
(1.631 0.9941 2 0.644 0.9522 1 6.936e 2 4 un(1) 1 5.991e 2 4 un
(2)))),[13.6 11.5]);
The initial conditions u(k 2 1) 5 13.6 at t 5 0.46 and u(k 2 2) 5 11.5 at t 5 0.42 are
obtained from the values of the control variable at time t 5 0.42 and t 5 0.46 with the initial
12.5 Sampling period switching
14
13.5
Control Signal
13
12.5
12
11.5
11
10.5
10
0.42
0.43
0.44
0.45
0.46 0.47
Time s
0.48
0.49
0.5
FIGURE 12.25
Control values for the two controllers described in Example 12.12. The solid line
represents the control variable with the faster controller. The dashed line represents the
equivalent control variable with the slower controller.
1
Process Output
0.8
0.6
0.4
0.2
0
0
0.5
1
1.5
Time s
2
2.5
FIGURE 12.26
Step response described in Example 12.12 with sampling period switching.
3
525
526
CHAPTER 12 Practical Issues
sampling period T0 5 0.01 s (Figure 12.25). The optimization yields the control values
u(0.46) 5 11.3995 and u(0.42) 5 12.4069. The resulting value of the objective function is
zero (i.e., the measured output is equal to the model estimate at the switching time).
Thus, the value of the control variable at the switching time t 5 0.5 s is
uð0:5Þ 5 11:3995 1 48:14 3 2 0:0172 2 46:26 3 0:0151 5 9:8730
The step response of the system, shown in Figure 12.26, has a small overshoot and
time to first peak and a short settling time. The response is smooth and does not have a
discontinuity at the switching point.
12.5.2 Dual-rate control
In some industrial applications, samples of the process output are available at a
rate that is slower than the sampling rate of the controller. If performance
degrades significantly when the controller sampling rate is reduced to equal that
of the process output, a dual-rate control scheme can be implemented. The situation is depicted in Figure 12.27, where it is assumed that the slow sampling
period λT is a multiple of the fast sampling period T (i.e., λ is an integer). Thus,
the ADC operates at the slower sampling rate 1/(λT), whereas the controller and
the sample-and-hold operate at the faster sampling rate 1/T.
To achieve the performance obtained when the output is sampled at the fast
rate, a possible solution is to implement the so-called dual-rate inferential control scheme. It uses a fast-rate model of the process G^ ZAS ðzÞ to compute the missing output samples. The control scheme is shown in Figure 12.28, where a is an
availability parameter for the output measurement defined by
0; t 6¼ kλT
a5
1; t 5 kλT
The controller determines the values of the control variable using the measured output at t 5 kλT and using the estimated output when t 5 kT and t6¼kλT. In
the absence of disturbances and modeling errors, the dual-rate control scheme is
equivalent to the fast single-rate control scheme. Otherwise, the performance of
dual-rate control can deteriorate significantly. Because disturbances and modeling
errors are inevitable in practice, the results of this approach must be carefully
checked.
E(z)
R(z)
+
−
U(z)
C(z)
Y(s)
ZOH
G(s)
λT
FIGURE 12.27
Block diagram of dual-rate control. The controller and the ZOH operate with sampling
period T; the process output sampling period is λT.
12.5 Sampling period switching
R(z)
E(z)
+
−
U(z)
C(z)
ZOH
G(s)
ĜZAS(z)
λT
a
1−a
FIGURE 12.28
Block diagram of dual-rate inferential control.
EXAMPLE 12.13
Design a dual-rate inferential control scheme with T 5 0.02 and λ 5 5 for the process (see
Example 6.16)
GðsÞ 5
1
ðs 1 1Þðs 1 10Þ
and the controller
CðsÞ 5 47:2
s11
s
Solution
The fast rate model (T 5 0.02) for the plant with DAC and ADC is
GZAS ðzÞ 5 1024
1:86z 1 1:729
YðzÞ
5
5 G^ ZAS ðzÞ
z2 2 1:799z 1 0:8025 UðzÞ
The difference equation governing the estimates of the output is
yðkÞ 5 1:799yðk 2 1Þ 2 0:8025yðk 2 2Þ 1 1:86 3 1024 uðk 2 1Þ 1 1:729 3 1024 uðk 2 2Þ
Process Output
1
0.8
0.6
0.4
0.2
0
0
0.5
1
1.5
2
2.5 3
Time s
3.5
4
4.5
5
FIGURE 12.29
Step response described in Example 12.13 for a dual-rate controller with T 5 0.02 and λ 5 5.
527
CHAPTER 12 Practical Issues
1.6
1.4
1.2
Process Output
528
1
0.8
0.6
0.4
0.2
0
0
0.5
1
1.5
2
2.5 3
Time s
3.5
4
4.5
5
FIGURE 12.30
Step response described in Example 12.13 for a single-rate controller with T 5 0.1.
With T 5 0.02, the controller transfer function obtained by bilinear transformation is
CðzÞ 5
47:67z 2 46:73
z21
The controller determines the values of the control variable from the measured output
at t 5 5kT. When t 5 kT and t 6¼ 5kT, we calculate the output estimates using the estimator
difference equation.
The control scheme shown in Figure 12.28 yields the step response shown in
Figure 12.29. The results obtained using a single-rate control scheme with T 5 0.1 are
shown in Figure 12.30. The step response for single-rate control has a much larger first
peak and settling time.
Resources
Albertos, P., Vallés, M., Valera, A., 2003. Controller transfer under sampling rate dynamic
changes. Proceedings European Control Conference. Cambridge, UK.
Åström, K.J., Hägglund, T., 2006. Advanced PID Controllers. ISA Press, Research
Triangle Park, NJ.
Cervin, A., Henriksson, D., Lincoln, B., Eker, J., Årzén, K.-E., 2003. How does control
timing affect performance? IEEE Control Syst. Mag. 23, 1630.
Gambier, A., 2004. Real-time control systems: A tutorial. Proceedings 5th Asian Control
Conference, pp. 10241031.
Li, D.S., Shah, L., Chen, T., 2003. Analysis of dual-rate inferential control systems.
Automatica 38, 10531059.
Visioli, A., 2006. Practical PID Control. Springer, London, UK.
Problems
PROBLEMS
12.1
Write pseudocode that implements the following controller:
CðzÞ 5
UðzÞ 2:01z 2 1:99
5
EðzÞ
z21
12.2
Rewrite the pseudocode for the controller described in Problem 12.1 to
decrease the execution time by assigning priorities to computational
tasks.
12.3
Design an antialiasing filter for the position control system
GðsÞ 5
1
sðs 1 10Þ
with the analog controller (see Example 5.6)
CðsÞ 5 50
s 1 0:5
s
Select an appropriate sampling frequency and discretize the controller.
12.4
Determine the mean and variance of the quantization noise when a 12-bit
ADC is used to sample a variable in a range of 0 to 10 V for (a) rounding
and (b) truncation.
12.5
For the system and the controller described in Problem 12.3 with a sampling interval T 5 0.02 s, determine the decrease in the phase margin due
to the presence of the ZOH.
12.6
Consider an oven control system (Visioli, 2006) with transfer function
GðsÞ 5
1:1
e225s
1300s 1 1
and the PI controller
CðsÞ 5 13
200s 1 1
200s
Let both the actuator and the sensor signals be in the range 0 to 5 V,
and let 1 C of the temperature variable correspond to 0.02 V. Design the
hardware and software architecture of the digital control system.
529
530
CHAPTER 12 Practical Issues
12.7
Write the difference equations for the controller in (a) direct form,
(b) parallel form, and (c) cascade form.
CðzÞ 5 50
12.8
ðz 2 0:9879Þðz 2 0:9856Þ
ðz 2 1Þðz 2 0:45Þ
For the PID controller that results by applying the Ziegler-Nichols tuning
rules to the process
1
e22s
GðsÞ 5
8s 1 1
determine the discretized PID controller transfer functions (12.11) and
(12.12) with N 5 10 and T 5 0.1.
12.9
Design a bumpless manual/automatic mode scheme for the PID controller
(T 5 0.1)
CðzÞ 5
12.10
Design a bumpless manual/automatic mode scheme for the controller
obtained in Example 6.19
CðzÞ 5
12.11
252z2 2 493:4z 1 241:6
ðz 2 1Þðz 2 0:13Þ
1:3932ðz 2 0:8187Þðz 2 0:9802Þðz 1 1Þ
ðz 2 1Þðz 1 0:9293Þðz 2 0:96Þ
Determine the digital PID controller (with T 5 0.1) in incremental form
for the analog PID controller
1
1 2s
CðsÞ 5 3 1 1
8s
COMPUTER EXERCISES
12.12
Write a MATLAB script and design a Simulink diagram that implements
the solution to Problem 12.8 with different filter parameter values N, and
discuss the set-point step responses obtained by considering the effect of
measurement noise on the process output.
12.13
Consider the analog process
GðsÞ 5
1
e22s
8s 1 1
Computer Exercises
and the analog PI controller with Kp 5 3 and Ti 5 8. Obtain the set-point
step response with a saturation limit of umin 5 21.1 and umax 5 1.1 and
with a digital PI controller (T 5 0.1) with
(a) No anti-windup
(b) A conditional integration anti-windup strategy
(c) A back-calculation anti-windup strategy
(d) A digital PI controller in incremental form
12.14
Consider the analog process and the PI controller described in Problem
12.13. Design a scheme that provides a bumpless transfer between manual and automatic mode, and simulate it by applying a step set-point signal and by switching from manual mode, where the control variable is
equal to one, to automatic mode at time t 5 60 s. Compare the results
with those obtained without bumpless transfer.
12.15
Design and simulate a dual-rate differential control scheme with T 5 0.01
and λ 5 4 for the plant
GðsÞ 5
1
ðs 1 1Þðs 1 5Þ
and the analog PI controller (see Problem 5.7). Then apply the controller
to the process
G~ ðsÞ 5
1
ðs 1 1Þðs 1 5Þð0:1s 1 1Þ
to verify the robustness of the control system.
12.16
Consider the analog process and the analog PI controller described in
Problem 12.13. Write a MATLAB script that simulates the step response
with a digital controller when the sampling period switches at time
t 5 0.52 from T 5 0.04 to T 5 0.01.
12.17
Consider the analog process and the analog PI controller described in
Problem 12.13. Write a MATLAB script that simulates the step response
with a digital controller when the sampling period switches at time
t 5 0.52 from T 5 0.01 to T 5 0.04.
531
APPENDIX
I
Table of Laplace and
z-Transforms
No.
Continuous
Time
Laplace
Transform
Discrete Time
z-Transform
1
δ(t)
1
δ(k)
1
2
1(t)
1
s
1(k)
z
z21
3
t
1
s2
kT
zT
ðz21Þ2
4
t2
2!
s3
(kT)2
zðz 1 1ÞT 2
ðz21Þ3
5
t3
3!
s4
(kT)3
zðz2 1 4z 1 1ÞT 3
ðz21Þ4
6
e2αt
1
s1α
ak
7
1 2 e2αt
α
sðs 1 αÞ
1 2 ak
ð1 2 aÞz
ðz 2 1Þðz 2 aÞ
8
e2αt 2 e2βt
β 2α
ðs 1 αÞðs 1 βÞ
ak 2 bk
ða 2 bÞz
ðz 2 aÞðz 2 bÞ
9
te2αt
1
ðs1αÞ2
kTak
az T
ðz2aÞ2
10
sin(ωnt)
ωn
s2 1 ω2n
sin(ωnkT)
sinðωn TÞz
z2 2 2cosðωn TÞz 1 1
11
cos(ωnt)
s
s2 1 ω2n
cos(ωnkT)
z½z 2 cosðωn TÞ
z2 2 2cosðωn TÞz 1 1
12
e2ζωn t sinðωd tÞ
ωd
ðs1ζωn Þ2 1 ω2d
e2ζωn kT sinðωd kTÞ
z
z2a
e2ζωn T sinðωd TÞz
22ζωn T
d TÞz 1 e
z2 2 2e2ζωn T cosðω
The discrete time functions are generally sampled forms of the continuous time functions.
Sampling t gives kT, whose transform is obtained by multiplying the transform of k by T.
The function e2αkT is obtained by setting a 5 e2αT.
533
534
APPENDIX I Table of Laplace and z-Transforms
No.
13
Continuous
Time
2ζωn t
e
cosðωd tÞ
Laplace
Transform
Discrete Time
s 1 ζωn
ðs1ζωn Þ2 1 ω2d
e
2ζωn kT
cosðωd kTÞ
z-Transform
z½z 2 e2ζωn T cosðωd TÞ
22ζωn T
d TÞz 1 e
z2 2 2e2ζωn T cosðω
14
sinh(βt)
β
s2 2 β 2
sinh(βkT)
sinhðβTÞz
z2 2 2coshðβTÞz 1 1
15
cosh(βt)
s
s2 2 β 2
cosh(βkT)
z½z 2 coshðβTÞ
z2 2 2coshðβTÞz 1 1
INDEX
Index Terms
Links
Note: Page numbers followed by “f”, “t” and “b” refer to figures, tables and boxes, respectively.
A
abs command
144
520
Absolute stability
481
482b
contraction
482b
input-output stability
481
Absolutely summable impulse response
sequences
Accuracy, in digital control
96b
300
2
acker command
395
Ackermann’s formula
357
Actuators
298
360
2
furnace model
118
linear gain block
484
noise damage
507
Nyquist plot of
484f
saturation nonlinearity
509
tank control system
ADCs (analog-to-digital converters)
models
504b
2
55
55
quantization errors
498
selecting
494
stable range
107
transfer functions
61
Addition of matrices
538b
Adjoint matrices
Aircraft turbojet engine
545
4
This page has been reformatted by Knovel to provide easier navigation.
Index Terms
Algebraic loops
Links
513
in controller structure
513
Algebraic Lyapunov equation
420
Algebraic multiplicity of eigenvalues
548
Aliasing
495
Allocation, pole
353
Analog control system design
127
computer exercises
503
163b
design specifications and gain variation
132
objectives
127
problems
161b
root locus method, see Root locus method
Analog controller design implementation
180
bilinear transformation
186
differencing methods
181
PID controller tuning
199
Analog disturbances
74
Analog filters
bilinear transformations
186
187f
differencing methods
181
181b
188b
Analog plants
cruise control systems
107
direct control design
216
PID controllers
199
position control systems
108
121
193b
203
speed control systems
190
195b
202b
209b
215b
219
232
217
Analog systems
inputs and outputs
1
piecewise constant inputs
9
stability
463
steady-state error
79b
This page has been reformatted by Knovel to provide easier navigation.
Index Terms
Links
Analog systems (Cont.)
transfer functions
61
Analog-to-digital converters, see ADCs
Analytic continuation
14
angle command
144
Angle of arrival, in root locus method
128
Angle of departure, on root locus method
128
Angles
bilinear transformations
170f
contours
109
root locus method
128
130
Anthropomorphic manipulators
state–space equations
241
Anthropomorphic manipulators
torque vectors
441b
Antialiasing filters
495
Anti-windup techniques
509
498
503
384b
387b
504
Armature-controlled DC motors
full-order observers
gain vector
375b
360
PID controllers
153b
reduced-order observers
379b
root locus method
137b
state equations
248
z-domain transfer function
67b
386b
Assignment, pole
353
Asymptote angles, in root locus method
130
Asymptotic stability
91b
295b
300b
301b
with BIBO stability
conditions
equilibrium points
global
355b
389
94
294b
452
453
This page has been reformatted by Knovel to provide easier navigation.
393
Index Terms
Links
Asymptotic stability (Cont.)
LTI systems
94b
with Lyapunov equations
454
open-loop stable affine system
473
Automatic mode for PID controllers
512
Autonomous underwater vehicles (AUVs)
287
97
348
B
Back-calculation technique
509
Backward differencing methods
182
Backward iteration, in Leverrier algorithm
252
Band-limited signals
45
Bandwidth
filters
496
sampling selection
46
second-order systems
47
Bessel filters
507
496
BIBO (Bounded-Input–Bounded-Output)
stability
94
Bilinear transformations
297
523
analog controller design implementation
186
frequency response design
206
Routh-Hurwitz criterion
102
Binding, in drug-delivery system
444
bode command
294
44
116
Bode plots
bilinear transformations
188
frequency response design
206
208b
212
212f
116
120f
92
92f
Nyquist criterion
Bounded sequences
209f
This page has been reformatted by Knovel to provide easier navigation.
210f
Index Terms
Links
Bounded-Input–Bounded-Output, see BIBO
stability
Breakaway points, in root locus method
128
Break-in points, in root locus method
128
Bumpless transfer, of PID controller modes
512
Butterworth antialiasing filters
504
C
c2d command
30
bilinear transformations
188
state–space equations
269
transfer functions
CAD (computer-aided design)
80
322
29
canon command
281
Canonical realization
326
Cascades
controller structure
507
PI controllers
147
root locus method
136f
transfer function
case statements
138
58
492
Causal signals
delay period
12
z-transforms
12
Causal system impulse response
16
33
Cayley-Hamilton theorem
controllability conditions
303
pole placement
357
Center equilibrium points
465
304
469f
Characteristic equations
bilinear transformations
191
controller structure
504
This page has been reformatted by Knovel to provide easier navigation.
Index Terms
Links
Characteristic equations (Cont.)
pole assignment
389
root locus method
127
similar systems
282
state plane analysis
465
Characteristic polynomial, z-domain
172
Chemical reactor
100
Circle criterion
input-output stability
Closed-loop drug-delivery system
Closed-loop eigenvalues
131
482b
482b
3
359
Closed-loop optimal control state matrix,
eigenvalues of
430
Closed-loop system
input-output stability
481f
linear and nonlinear blocks
481f
Closed-loop transfer functions
52b
71
73
Coefficients
MATLAB
30
partial fraction expansion
24
pole placement by
355b
Cofactors of matrices
544
Column vectors
537
539b
Columns of matrices
rank
547
transposing
538
Completely controllable systems
302
Completely observable systems
313
Complex conjugate eigenvalues
discrete-time state–space equations
269
real form for
262
This page has been reformatted by Knovel to provide easier navigation.
Index Terms
Links
Complex conjugate poles
partial fraction expansion
24
root locus method
135
z-domain
168
169t
Complex differentiation property
18
535t
Complex integrals
18
Computed torque method
173
441
Computer-aided design, see CAD
Conditional integration
Constant inputs, in analog systems
Constant vectors, in linear equations
Constants, error
Constituent matrices
509
9
242
75
255
properties
260
state matrices
258
Contours
Nyquist criterion
109
z-domain
171
Contraction constants
449
Contraction functions
449
Contraction
absolute stability
482b
sector bound nonlinearity
481b
Control effort, in pole placement
359
Control jitter
491
Control matrices
239
Control signals, in sampling period switching
521f
Controllability
301
matrices
304
in normal form
308
Controllable canonical realization
326
This page has been reformatted by Knovel to provide easier navigation.
Index Terms
Links
Controllable form
MATLAB
330
transformation to
356
Controllable systems
302
Controller form, of Lyapunov equation
457
Controller structure
504
algebraic loop in
513
513
Convergence rate, in Lyapunov stability
theory
454
Conversion delay, in ADCs
494
Convolution summation
BIBO stability
95
cascade transfer function
59
discrete-time state equations
discrete-time systems
273
32
Convolution theorem
discrete-time systems
34
sampling rate selection
46
Coprime transfer functions
389
Cost function
399
Cost of digital controls
2
Costate equations
406
CPU utilization optimization
516
Critical gain, in z-domain root locus
166b
167b
Crossover frequency, in analog
control systems
503
Cruise control systems
frequency response design
208b
impulse response
63b
stable range
107
ctrb command
307
Cubic interpolation
522
This page has been reformatted by Knovel to provide easier navigation.
Index Terms
Cut-off frequency, in low-pass filters
Links
496
D
DACs (digital-to-analog converters)
models
2
55
56
selecting
494
stable range
107
transfer functions
61
498
499
Damping ratio
analog control system design
sampling frequency
stability determination
133
48
101
dare command
421
Data acquisition systems
494
DC gain
in frequency response
proportional control
42
175
DC motor position control systems
bilinear transformations
193b
root locus method
137b
z-domain digital control system design
203
DC motor speed control systems
bilinear transformations
190b
deadbeat controllers
219b
direct control design
215b
frequency response design
209b
sampling period switching
523
servo problem
368b
z-domain digital control system design
202b
225b
DC motors
controllability
306b
full-order observers
375b
384b
387b
This page has been reformatted by Knovel to provide easier navigation.
Index Terms
Links
DC motors (Cont.)
PID controllers
153b
reduced-order observers
379b
z-domain transfer function
ddamp command
386b
67b
101
Deadbeat controllers
gain and zero state
363
isothermal chemical reactor
228b
speed control system
219b
vehicle positioning system
224b
225b
Decaying exponential sequences, in
z-transforms
Decoupling zeros
29
278
320
Delay
ADC conversion
494
antialiasing filters
497
controllable canonical realization
329
sampling rate selection
47
sensor
47
transport
69
zero-order hold (ZOH)
57
503
z-transforms
16
37
den command
30
denp command
30
Dependent vectors
547
Derivative action filtering
507
Derivative time constant
156
Derivatives of matrices
561
det command
544
Detectability
317
69
535t
338
Determinants
partitioned matrices
551
This page has been reformatted by Knovel to provide easier navigation.
Index Terms
Links
Determinants (Cont.)
square matrices
Deterministic disturbances
544
74
Diagonal form, for state–space equations
281
Diagonal matrices
257
543b
Difference equations
canonical realization
326
discrete-time systems
11
Lyapunov
412
nonlinear
447
z-transforms
Differences, of partitioned matrices
31
551
Differencing methods, in analog controller
design
181
Differentiation, of matrices
Digital control overview
561
1
advantages
2
examples
3
objectives
1
problems
7b
structure
2
Digital control systems design
computer exercises
165
233b
objectives
165
problems
230b
z-domain design, see Z-domain digital
control system design
z-domain root locus
Digital control systems modeling
165
55
ADC models
55
analog disturbances
74
closed-loop transfer function
71
This page has been reformatted by Knovel to provide easier navigation.
Index Terms
Links
Digital control systems modeling (Cont.)
combinations transfer function
computer exercises
61
89b
DAC models
56
MATLAB commands
79
objectives
55
problems
85b
sampler effect on cascade transfer function
58
steady-state error and error constants
75
transport lag
69
ZOH transfer function
57
Digital filters
bilinear transformations
186
differencing methods
182
Digital signal processing (DSP) chips
187f
188b
4
Digital-to-analog converters, see DACs
Diophantine equation
Dirac deltas
389
392
39
Direct control design
200
Direct controller form
507
Direct transmission matrices
239
Discrete filter blocks
489
Discrete Lyapunov equations
456
Discrete transfer functions
206
Discrete-time convolution property
535t
Discrete-time frequency response
Discrete-time nonlinear controller design
44
470
extended linearization
470
Lyapunov stability theory
470
Discrete-time state–space equations
Discrete-time systems
analog systems with piecewise constant inputs
213
266
277
9
9
This page has been reformatted by Knovel to provide easier navigation.
Index Terms
Links
Discrete-time systems (Cont.)
CAD
29
computer exercises
52b
difference equations
11
frequency response
39
objectives
problems
9
49b
sampling theorem
45
time response
32
z-transforms, see Z-transforms
Discretization of nonlinear systems
439
by input and state redefinition
442
by input redefinition
440
by matching conditions
445
by output differentiation
443
Discretized controller transfer function
508
Distinct eigenvalues
548
Distortion
analog filters
190
approximation
200
frequency response
Disturbances, analog
43
186
206
423
425
74
dlqr command
421
dlqry command
421
dlyap command
457
Domain of attraction
461
Double integrator systems
Hamiltonian system
427
mechanical systems
412b
steady-state regulator for
419
Drug-delivery system
drug-receptor binding in
444
This page has been reformatted by Knovel to provide easier navigation.
208
Index Terms
Links
Drug-delivery system (Cont.)
example
3
dsort command
DSP chips
101
4
Duality
338
Dual-rate inferential control scheme
526
Duals
338
E
Effective bandwidth, in sampling selection
47
495
eig command
261
281
Eigenpairs
353
Eigenstructure of matrices
258
Eigenstructure, of Hamiltonian matrix
429
Eigenstructure, of inverse Hamiltonian system
352
432b
Eigenvalues, complex conjugate
discrete-time state–space equations
269
real form for
262
Eigenvalues
matrices
548
observer state feedback
383
388
pole placement
355
359
state–space equations
281
Eigenvectors
548
Engines, turbojet
Equal matrices
4
538b
Equating coefficients
partial fraction expansion
pole placement by
proportional control systems
24
355b
137
140
294b
452
150
Equilibrium points
asymptotic stability
This page has been reformatted by Knovel to provide easier navigation.
156
Index Terms
Links
Equilibrium points (Cont.)
classification
465t
linearized models
244
nonlinear discrete-time systems
448
nonlinear spring-mass-damper systems
245
state plane analyses
465
unstable
459
Equilibrium state of systems
294
Equivalent norms
553
Equivalent systems
282
Errors
controller
504
integrator windup
509
quantization
494
sampling period switching
521f
state estimators
374
498
steady-state, see Steady-state errors
tracking
75
Estimating domain of attraction
461
Estimation error
374
evalfr command
143
Execution time reductions
493
expm command
269
Exponential change, in root locus method
135
Exponential stability
463
Exponential z-transforms
15b
Extended linearization
440
discrete-time nonlinear controller design
470
by input and state redefinition
442
by input redefinition
440
by matching conditions
445
by output differentiation
443
17
29
447
This page has been reformatted by Knovel to provide easier navigation.
38b
Index Terms
Links
F
Feedback
compensation, in root locus method
136f
139
inertial control system
414f
415f
parallel realization
331
stability determination
101
139f
state, see State feedback control
Feedforward action
370
Filtering observers
376
Filters
antialiasing
496
503
bilinear transformations
186
187f
derivative actions
507
differencing methods
181
Final value theorem
28
Finite impulse response (FIR) systems
97
Finite settling time design
218
First-order approximation of linear equations
243
First-order holds
188b
535t
56
First-order systems
frequency response
47
optimal control
407
root locus
166f
First-order-plus-dead-time model
158t
Fixed point equilibrium
448
Flexibility, in digital controls
2
Fluid level control system
10f
fminsearch command
520
Folding
43
frequency
Forward differencing methods
43
181
This page has been reformatted by Knovel to provide easier navigation.
Index Terms
Links
Fourier transform
45b
Free final state, in linear quadratic regulators
410
Frequency of oscillations, in z-domain function
174
Frequency response
discrete-time systems
39
distortion
206
z-domain digital control system design
205
ZOH
57
Frobenius norms
554
Full rank matrices
547
Full-order observers
374
503
384b
387b
Furnace
closed loop stability
113
linear gain block
484
Nyquist criterion
118
Nyquist plot of
484f
z-domain transfer function
66b
G
Gain margin
frequency response design
205
Nyquist criterion
114
Gain matrices, for full-order observers
375b
Gain vector, in Ackermann’s formula
360
361
analog control systems
132
503
bilinear transformations
191
192
deadbeat controllers
363
Gain
discrete-time systems frequency response
42
inertial control system
414f
PI controllers
151
z-domain root locus
415f
166b
This page has been reformatted by Knovel to provide easier navigation.
Index Terms
Links
Global asymptotic stability
453
Global linearization
440
Globally positive definite functions
450
Gradient matrices
562
Gradient of inner product
562
Gradient vectors
562
H
Hamiltonian function
linear quadratic regulators
410
optimal control
404
Hamiltonian system
426
eigenstructure of
429
Hankel realization
339
Hardware design
494
Hermitian matrices
Hessian matrices
Homogeneous equations
406
539b
401
407
562
11
I
Idempotent matrices
Identity matrices
If-then-else statements
Implementation error reductions
260
543b
492
2
Impulse disturbances
74
Impulse response sequence
32
367
Impulse response
BIBO stability
95
cascaded
58
convolution theorem
35
DAC transfer functions
62
298b
299b
63f
This page has been reformatted by Knovel to provide easier navigation.
Index Terms
Links
Impulse response (Cont.)
discrete-time systems
discrete-time state–space equations
32
277
FIR
97
Laplace transforms
57
second-order systems
48
z-transforms
75
264
277
Impulse sampling
discrete-time waveforms
39
Laplace transforms
74
LTI systems
59
sampling theorem
46
Incremental form, for PID controllers
515
Indefinite functions
451
Indefinite quadratic forms
557
Independent vectors
547
Induced matrix norms
554
Inertial control systems
feedback and gain
415f
phase plane trajectory
415f
position trajectory
414f
steady-state regulator
421
velocity trajectory
414f
INFANTE underwater vehicle
287
Initial command
237
Initial value theorem
535t
Initialization step, in Leverrier algorithm
252
Inner loop feedback compensation
136f
Input decoupling zeros
320
Input differencing
326
Input matrices
239
Input redefinition, extended linearization by
440
425
348
This page has been reformatted by Knovel to provide easier navigation.
Index Terms
Links
Input zero direction
322
Input-decoupling zeros
278
Input-output stability
474
absolute stability
481
circle criterion
482b
definition
475b
proof
479b
small gain theorem
478b
Input–output-decoupling zeros
278
Instability theorems
459
Integral control, in servo problem
370
Integral controllers
147
Integral time constant, in PID controllers
156
Integration of matrices
561
Integrator windup
509
Internal dynamics
444
Internal stability
320
370f
371b
98
interp1 command
519
Interpolators
518
Intersample oscillations
219
522
222
359
Invariance
system zeros
372
transfer function
282
Invariant zeros
Inverse Hamiltonian system, eigenstructure of
324
432b
Inverse matrices
derivatives
561b
partitioned
551
square
545
Inverse transform
discrete-time state equations
z-transform
273
19
This page has been reformatted by Knovel to provide easier navigation.
463
Index Terms
Links
Ip norms
553
Irreducible state–space realization
278
Isothermal chemical reactor
100
deadbeat controller
228b
iztrans command
30
J
Jacobians
245
Jitter
491
Joseph form, of Riccati equation
412
Jury stability test
104
562
420
K
kron command
458
563
Kronecker product
457
563
402
407
L
Lagrange multiplier
of Hamiltonian matrix
429
Laplace transforms
cascade transfer function
58
closed-loop transfer function
72b
deadbeat controllers
221
differencing methods
181
discrete-time state equations
274
frequency response
furnace system
impulse-sampled output
39
66b
74
Leverrier algorithm
253
linear state–space equations
246
root locus method
135
This page has been reformatted by Knovel to provide easier navigation.
Index Terms
Links
Laplace transforms (Cont.)
tables of
533t
transfer function matrices
264
transport delay transfer function
z-domain function
69
169t
ZOH transfer function
57
z-transforms
12
Least-squares estimates
170
66b
401
Left half plane (LHP)
analog filters
180
bilinear transformations
189
PID controllers
153
root locus method
135
Routh-Hurwitz criterion
102
z-domain controllers
201
Left inverse
560
Length of eigenvectors
550
206
Leverrier algorithm
discrete-time state equations
272
linear state–space equations
251
pole placement
356
transfer function matrix
264
Linear difference equations
31b
Linear quadratic regulators
409
Linear quadratic tracking controllers
423
Linear resistance least-squares estimates
401
Linear state–space equations
complex conjugate eigenvalues, real form
for
262
Leverrier algorithm
251
vs. nonlinear
240
solution
246
This page has been reformatted by Knovel to provide easier navigation.
Index Terms
Links
Linear state–space equations (Cont.)
state-transition matrices
257
Sylvester’s expansion
255
Linear time-invariant (LTI) systems
asymptotically stable
94b
convolution summation
32
with impulse-sampled input
59
state-transition matrices
Linearity, of z-transforms
97
248
16
535t
Linearization
Lyapunov method
458
nonlinear state equations
243
Linearly dependent vectors
547
Linearly independent vectors
547
linsolve command
392
Local maximum
401
Local minimum
401
Locally positive definite functions
451
Logarithmic transformations
448
Long division, for inverse z-transforms
520
19
Loop gains
bilinear transformations
191
contours
109
frequency response design
205
internal stability
99b
100b
Nyquist criterion
109
113b
PD controllers
138
143
root locus method
127
steady-state error
75
z-domain controllers
204
Loops
algebraic
513
This page has been reformatted by Knovel to provide easier navigation.
Index Terms
Links
Loops (Cont.)
feedback
331
Lower triangular matrices
549
Low-pass filters
496
lqr command
421
lsim command
237
507
LTI, see Linear time-invariant systems
Lyapunov equations
algebraic
420
difference
412
Lyapunov stability theory
450
controller design based
473
convergence rate
454
discrete-time nonlinear controllers
470
domain of attraction estimation
461
functions
450
instability theorems
459
linear systems
454
linearization method
458
MATLAB
457
theorems
452
M
MACSYMA package
30
Manipulators
example
4
state–space equations
torque vectors
Manual mode, for PID controllers
240
242f
441b
512
MAPLE package
30
margin command
116
Marginally stable systems
91b
294
This page has been reformatted by Knovel to provide easier navigation.
Index Terms
Markov parameter matrices
Links
339
Matching conditions, extended linearization
using
445
MATHEMATICA package
30
MATLAB program
bilinear transformations
CAD
188
194
29
30
contours
174
controllability testing
307
controllable form
330
frequency response
44
gain crossover frequency
504
Hamiltonian systems
428
linear quadratic regulators
412b
Lyapunov equation
457
nonlinear controller design
470
Nyquist criterion
116
observability
316
PD controllers
143
PI controllers
151
PID controllers
154
point mass motion
237
pole assignment
366
pole defining
320
pole placement
364
position control systems
120f
392
79
root locus method
129
sampling period switching
519
stability determination
101
state estimators
375
state–space equations
269
state–space representation
240
138
281
This page has been reformatted by Knovel to provide easier navigation.
Index Terms
Links
MATLAB program (Cont.)
state-transition matrices
261
steady-state quadratic regulators
421
transfer function matrices
265
z-domain control system design
175
z-domain root locus
166b
zeros
323b
z-transfer function
425
201
202
279
Matrices
addition and subtraction
538
column and row vectors
539b
constituent
255
controllability
304
diagonal state
257
differentiation and integration
561
discrete-time state equations
274
eigenstructure
352
eigenvalues
548
eigenvectors
548
equal
538b
gain
375b
impulse response
300
inverse
545
Kronecker product
457
Leverrier algorithm
251
linear state–space equations
247
multiplication
540
norms
554
orthogonal
546
quadratic forms
555
rank
547
representation
258
563
547b
537b
This page has been reformatted by Knovel to provide easier navigation.
Index Terms
Links
Matrices (Cont.)
state
257
state–space representation
240
traces
546
transfer function
264
transposition
538
z-transform solution
272
547b
Matrix exponential
discrete-time state–space equations
267
Leverrier algorithm for
253
linear state–space equations
247
of state matrices
259
Matrix inversion lemma
411
Matrix polynomials, for pole placement
357
Matrix Riccati equation
411
Maximization, of functions
399
Mechanical systems
Milling machine
552
412b
322
MIMO systems, see Multi-input–multi-output
(MIMO) systems
Minimal polynomials
320
Minimal state–space realization
278
Minimization of functions
399
Minimum control effort
407
Minimum phase
bilinear transformations
196
frequency response
205
Minimum principle of Pontryagin
optimal control
406
linear quadratic regulators
410
Minors of matrices
544
Modal matrices
281
This page has been reformatted by Knovel to provide easier navigation.
Index Terms
Links
Modes
system
248
unobservable
313
Modified z-transform
37
Monic transfer functions
389
Motion, point mass
237
Motors, see DC motors
Multi-input systems, pole placement for
364
Multi-input–multi-output (MIMO) systems
data acquisition systems for
494
feedback gains
363
linear time-invariant systems
239
parallel realization for
334
poles and zeros
319
representation
235
transfer function matrices
264
Multiplexers (MUX)
Multiplication by exponential property
494
17
535t
540
551
Multiplication
matrices
polynomials
30
Multiplicity of eigenvalues
548
Multivariable systems
319
N
Nanopositioning
126
Necessity proofs
asymptotic stability
296
455
BIBO stability
96b
299
controllability
303
308
internal stability
99b
observability
315
This page has been reformatted by Knovel to provide easier navigation.
Index Terms
Links
Necessity proofs (Cont.)
state feedback
Negative definite functions
Negative frequencies transfer function
Negative semidefinite functions
353
451
556
42
451
Newton’s law, in cruise control system
63
nichols command
44
557
Noise
ADC resolution
494
antialiasing filters
496
observers
377
observers
383
PID derivative actions
507
quantization
498
Noncummutative matrices
540
Nonlinear closed-loop system
478f
Nonlinear difference equations
448
Nonlinear digital control systems
439
computer exercises
489b
difference equations
447
discrete-time
470
discretization
439
equilibrium of discrete-time system
448
equilibrium states
294
input-output stability
474
504
485
Lyapunov stability theory, see Lyapunov
stability theory
objectives
439
problems
485b
stability of analog systems with digital
control
state plane analysis
463
465
This page has been reformatted by Knovel to provide easier navigation.
Index Terms
Links
Nonlinear saturation block
475
Nonlinear spring-mass-damper system
245
Nonlinear state–space equations
vs. linear
240
linearization
243
Nonminimal state–space realization
278
Nonminimum phase system
206
Nonsingular matrices
544
Norm axioms
553
Normal form systems
controllability in
observability
308b
317
Normal matrix multiplication
540
Normalized eigenvectors
550
550
Norms
matrices
554
vectors
552
num command
30
Numerical controller errors
nyquist command
504
44
116
phase margin and gain margin
114
123
small gain theorem
479
stability
109
theorem
113b
Nyquist criterion
Nyquist frequency
495
504
O
Observability matrix
Observability rank condition
315
315b
Observability
313
Observable realization
336
This page has been reformatted by Knovel to provide easier navigation.
Index Terms
Observable systems
Links
313
Observer form
Lyapunov equation
457
Observer realization
simulation diagrams
337f
Observer state feedback
380
eigenvalue choices
383
Observers
full-order
374
384b
reduced-order
377
386b
obsv command
387b
316
Open-loop poles
Nyquist criterion
109
root locus method
127
111
Open-loop stable systems
asymptotic stability
473
optimal control
408
Open-loop state estimators
374
Optimal control
399
computer exercises
436b
constrained optimization
402
Hamiltonian systems
426
linear quadratic regulators
409
objectives
399
performance measures
404
problems
433b
steady-state quadratic regulators
419
unconstrained optimization
400
Optimal feedback gain matrix
in eigenstructure of inverse
Hamiltonian system
432
This page has been reformatted by Knovel to provide easier navigation.
Index Terms
Optimization
Links
399
constrained
402
CPU utilization
516
unconstrained
400
Origin
asymptotic stability
294
stabilizing
474
unstable equilibrium
459
Orthogonal matrices
546
Oscillations
closed-loop eigenvalues
359
conjugate poles
168
deadbeat controllers
222
final value theorem
28
finite settling time
218
gain variation
132
intersample
219
root locus design
135
z-domain function
174
Output differentiation
443
363
359
222
359
Output equations, in state–space
representation
238
Output feedback, in pole placement
367
Output matrices
239
Output quadratic regulators
420
Output zero direction
322
Output-decoupling zeros
278
Oven control systems
120f
320
Overshoot
analog control systems
132
frequency response design
211
This page has been reformatted by Knovel to provide easier navigation.
463
Index Terms
Links
P
Parallel controller form
507
Parallel realization
331
for multi-input-multi-output systems
334
Parameter perturbations
505
Parseval’s identity
477
Partial fraction coefficients
frequency response
41
MATLAB
30
Partial fraction expansion
20
Partitioned matrix multiplication
550
PD (proportional-derivative) controllers
138
bilinear transformations
28
189
195
pdcon function
145
154
Peak time, in analog control system design
132
Percentage overshoot, in analog control system
design
132
Performance measures
400
Performance
399
see also Optimal control
ADC resolution
494
linear quadratic regulators
409
Periodic nature, in frequency response
42
Perturbations
controller structure
504
linear equations
243
Nyquist criterion
115
tank control systems
507
119
9
Phase delay
antialiasing filters
497
zero-order hold (ZOH)
503
This page has been reformatted by Knovel to provide easier navigation.
Index Terms
Links
Phase margin
frequency response design
205
Nyquist criterion
114
zero-order hold (ZOH)
503
Phase plane
inertial systems
415f
state variables
236
Phase portraits
236
Phase shift, in ADCs
494
Phase variables
236
PI (proportional-integral) controllers
147
bilinear transformations
189
tank control system
503
422f
326
195
PID (proportional-integral-derivative)
controllers
153
507
bilinear transformations
190
195
bumpless mode transfer
512
derivative actions filtering
507
incremental form
515
integrator windup
509
tuning
156
Piecewise constant inputs
198f
395
470
199
9
place command
364
375
Planning horizons
419
428t
Plants
197f
9
plot command
44
Point mass
motion
237
state–space equations
280
pole command
320
Pole placement
353
closed-loop eigenvalues
359
This page has been reformatted by Knovel to provide easier navigation.
Index Terms
Links
Pole placement (Cont.)
eigenvalues
by equating coefficients
355
355b
MATLAB commands for
364
matrix polynomials for
357
for multi-input systems
364
by output feedback
367
stability
93
by transformation to controllable form
Pole polynomials
356
320
Poles
BIBO stability
96b
bilinear transformations
190
internal stability
99b
multivariable systems
319
Nyquist criterion
109
partial fraction expansion
100b
24
PI controllers
147
root locus method
127
stability determination
102
transfer functions
320
389
z-domain function
168
170
135
173
Pole-zero cancellation
asymptotically stable
94b
bilinear transformations
190
direct control design
213
frequency response design
206
PD controllers
138
z-transfer function
278
Pole-zero matching
183
poly command
364
143
This page has been reformatted by Knovel to provide easier navigation.
Index Terms
Links
Polynomials
MATLAB
30
matrix
357
pole
320
stability test
106
z-domain
172
polyvalm command
364
Pontryagin minimum principle
optimal control
406
linear quadratic regulators
410
Position control systems
bilinear transformations
193b
deadbeat controllers
224b
gain variation effect
133b
Nyquist criterion
121
PD controllers
145b
PI controllers
150b
root locus method
137b
stable range
108
steady-state position error
78b
z-domain digital control systems
z-transfer function
177b
122f
203
278
Position trajectory, in inertial control system
414f
Positive definite functions
451
Positive feedback loops
331
Positive integral power, of square matrices
541
Positive semidefinite functions
451
Positive systems
286
Postmultiplication of matrices
540
Practical issues
491
computer exercises
530b
controller structure
504
556
557
This page has been reformatted by Knovel to provide easier navigation.
Index Terms
Links
Practical issues (Cont.)
hardware and software architecture design
491
objectives
491
PID controllers, see PID controllers
problems
529b
sampling period switching
516
sampling period
495
Prediction observers
376
Premultiplication of matrices
540
541
Prewarping equality
187
188f
Primary strips
170
170f
Product concentration control
100
Production level profit estimates
402b
Profit estimates
402b
Programming requirements
491
Proofs
asymptotic stability
296b
455b
BIBO stability
96b
299b
controllability
303b
308b
exponential stability
463b
input-output stability
479b
internal stability
473
99b
observability
314b
state feedback
353b
uncontrollable systems
310b
Properties
constituent matrices
260
frequency response
42
262
state–space models,
see State–space models properties
z-transforms
15
535t
This page has been reformatted by Knovel to provide easier navigation.
Index Terms
Links
Proportional control
root locus method
137
z-domain digital control system design
175
Proportional-derivative controllers, see PD
controllers
Proportional-integral controllers, see PI
controllers
Proportional-integral-derivative controllers,
see PID controllers
Pseudocode
493
Pseudoinverses
of full-rank matrix
560
singular value decomposition and
557
Q
Quadratic forms
Lyapunov functions
451
matrices
555
Quadratic regulators
linear
409
steady-state
419
Quadruples
240
Quantization errors
498
ADC models
55
effects of
498
word length effects
494
QUOTE
269
280
56
262
R
rank command
controllability tests
307
This page has been reformatted by Knovel to provide easier navigation.
Index Terms
Links
rank command (Cont.)
Lyapunov equation
458
observability tests
316
Ranks
controllability
matrices
304b
260
observability
547
547b
315b
Rate of convergence, in Lyapunov stability
theory
454
Rational loop gains
205
RC filters
496
Reachability
302
Real roots, in partial fraction expansion
21
24
Realizations, see State–space realizations
Real-time systems
492
Rectangular matrices
537
Reduced-order observers
377
Reducible state–space realization
278
Reducible transfer function
338
Relative degree
444
Relative stability
114
Repeated eigenvalues
548
Repeated roots, in partial fraction expansion
26
residue command
30
Residues, in partial fraction expansion
26
Resolution, in ADCs
386b
30
494
Resolvent matrices
discrete-time state equations
274
Leverrier algorithm
251
linear state–space equations
247
state matrices
257
transfer function matrices
266
This page has been reformatted by Knovel to provide easier navigation.
Index Terms
Links
Riccati equation
Hamiltonian systems
426
linear quadratic regulators
411
steady-state quadratic regulators
419
Right half plane (RHP)
bilinear transformations
206
closed-loop stability
114
root locus method
131
z-domain controllers
200
Right inverse
135
167
144
146
560
Ripple-free deadbeat controller
speed control system
225b
vehicle positioning system
224b
Rise time, in analog control system design
R-L circuits
132
65
rlocus command
132
175
Robotic manipulators
example
4
state–space equations
torque vectors
240
441b
Robustness, in closed-loop poles selection
363
Root locus method
135
analog control system design
127
PD controllers
138
PI controllers
147
PID controllers
153
proportional control
137
z-domain
165
roots command
101
Rosenbrock’s system matrix
323
Rotation, in Nyquist criterion
115
204
This page has been reformatted by Knovel to provide easier navigation.
165
Index Terms
Links
Rounding, in ADCs
498
Routh tables
130
Routh-Hurwitz criterion
102
Row vectors
537
539b
Rows of matrices
rank
547
transposing
538
S
Saddle points
465t
467f
Sampled parabolic input
75
78
Sampled ramp input
19
75
Sampled step input
14b
75
Samplers, in cascade transfer function
58
61
Sampling frequency selection
46
Sampling period switching
516
dual-rate control
526
MATLAB
519
Sampling periods
495
antialiasing filters
496
bilinear transformations
186
quantization errors
498
ZOH phase delay
503
Sampling theorem
45
Sampling, for ADC models
55
Saturation nonlinearity
476b
Scalar continuous functions
451
Scalar Lyapunov functions
451
Scalar systems, in optimal control
407
Scalars, matrix multiplication by
540
Scale factors, in frequency response
39
Scaled vectors, in Nyquist criterion
115
495
509
116
This page has been reformatted by Knovel to provide easier navigation.
Index Terms
Links
Schur stable systems
296
Schur-Cohn test
104
s-degree-of-freedom (s-D.O.F.) robotic
manipulator
240
s-domain poles
168
Second-order holds
Sector bound nonlinearity
contraction
Semidefinite functions
Sensors
171t
202
56
481b
481b
451
3
Sensors
aircraft engine
delay
4
47
drug-delivery system
noise
69
3
74
Separation theorem
381
Series R-L circuits
65
Servo problem
367
set command
240
494
372
Settling time
analog control system design
132
bilinear transformations
191
finite
218
frequency response design
211
z-domain function
174
Shannon reconstruction theorem
134
52
Signs
cofactors
544
quadratic forms
556
Similarity transformation
279
Simplex algorithm
519
SIMULINK
80
480
This page has been reformatted by Knovel to provide easier navigation.
Index Terms
Simultaneous sample-and-hold (SSH) systems
Single-axis milling machine
Links
494
322b
Single-input (SI) case
239
Single-input–single-output (SISO) systems
235
eigenvalues
363
linear time-varying system
239
poles and zeros
319
transfer function matrices
264
transfer functions
325
transformation to controllable form
356
Single-output (SO) case
239
Singular matrix determinants
544
Singular value decomposition
and pseudoinverses
Singular value
Sinusoidal input
557
461
39
SISO systems, see Single-input–single-output
systems
Small gain theorem
Input-output stability
478b
478b
Software requirements
491
Software verification
492
Spectral radius of matrices
548
Spectrum of matrices
548
Speed control systems
bilinear transformations
190b
195b
deadbeat controllers
219b
225b
direct control design
215b
frequency response design
209b
sampling period switching
523
servo problem
368b
z-domain digital control system design
202b
This page has been reformatted by Knovel to provide easier navigation.
Index Terms
Speed of digital controls
Links
2
Spring-mass-damper system
245
sqrt command
154
Square matrices
determinants
544
inverse
545
representation
537
ss command
240
ss2ss command
281
279
SSH, see Simultaneous sample-and-hold systems
Stability
analog systems with digital control
463
asymptotic
94
294
BIBO
94
294
computer exercises
297
125b
conditions
94
definitions of
91
determination
101
internal
98
Jury test
104
Lyapunov, see Lyapunov stability theory
Nyquist criterion
objectives
problems
109
91
123b
z-domain pole locations
93
Stabilizability
309
338
Stable focus equilibrium points
465t
467f
Stable node equilibrium points
466f
Standard discrete-time signals for z-transforms
13
Standard form
for unobservable systems
318b
for uncontrollable systems
310b
This page has been reformatted by Knovel to provide easier navigation.
Index Terms
Links
State equations
performance measures
406
state–space representation
238
State estimation
full-order observers
374
measurement vectors
313
reduced-order observers
377
State feedback control
351
computer exercises
397b
feedback
351
invariance of system zeros
372
objectives
351
observer state feedback
380
pole assignment
389
pole placement, see Pole placement
problems
393b
servo problem
367
state estimation
374
State plane analysis
465
State planes
236
State portraits
236
State redefinition
442
State trajectories
236
State variables
235
State vectors
236
State-feedback gain matrices
423
State–space equations
linear, see Linear state–space equations
nonlinear
240
tracking controller
423
State–space models properties
computer exercises
293
349b
This page has been reformatted by Knovel to provide easier navigation.
Index Terms
Links
State–space models properties (Cont.)
controllability
301
detectability
317
duality
338
objectives
293
observability
313
poles and zeros of multivariable systems
319
problems
344b
stability
294
stabilizability
309
State–space realizations
325
controllable canonical realization
326
MATLAB controllable form
330
observable form
336
parallel realization
331
stability
294
State–space representation
235
computer exercises
289b
discrete-time state–space equations
319
238
266
linear state–space equations, see Linear
state–space equations
linearization of nonlinear state equations
243
MATLAB
240
nonlinear state–space equations
240
objectives
235
problems
283b
similarity transformations
279
state variables
235
transfer function matrices
264
zeros from
323
z-transfer function from
277
This page has been reformatted by Knovel to provide easier navigation.
Index Terms
Links
State-transition matrices
for diagonal state matrices
257
discrete-time state–space equations
268
Hamiltonian systems
426
Leverrier algorithm
254
linear state–space equations
248
uncontrollability
z-transform solution
306b
271
309
272
Steady-state errors
bilinear transformations
189
digital control systems
75
direct control design
214
frequency response design
209
frequency response design
211
PD controllers
140
PID controllers
155
servo problem
367
steady-state quadratic regulators
424
z-domain digital control system design
204
Steady-state quadratic regulators
216
145
419
linear quadratic tracking controllers
423
MATLAB
421
output quadratic regulators
420
Steady-state regulator problem
419
425
Steady-state response
desirable
132
frequency response
39
impulse disturbances
75
PID controllers
153
type number for
147
step command
138
Step function
38
175
This page has been reformatted by Knovel to provide easier navigation.
Index Terms
Links
Step response
bilinear transformations
192f
193
198f
deadbeat controllers
219
222
222f
direct control design
218f
219
dual-rate inferential control
526
frequency response design
211f
full-order observers
388
integrator windup
509
invariance of system zeros
373
nonlinear controller design
470
pole assignment
393
root locus method
138
sampling period switching
520
servo problem
369
steady-state regulator
425f
z-domain digital control system design
205f
Structure, of digital controls
Submultiplicative property
subplot command
Subtraction of matrices
212
373f
523
2
554
44
538b
Sufficiency proofs
asymptotic stability
296
455
BIBO stability
96b
299
controllability
303
308
internal stability
99b
observability
314
state feedback
354
Sum of squares, for linear equations
243
Sums, of partitioned matrices
551
Switching, sampling period
516
dual-rate control
526
MATLAB
519
This page has been reformatted by Knovel to provide easier navigation.
Index Terms
Sylvester’s expansion
Links
255
Symbolic manipulation toolbox
30
Symmetric matrix transpositions
539
Symmetry, of frequency response
42
syms command
30
Synthesis approach
System state definition
297
213
236b
System trajectories
454
System transfer function
36b
System zeros
264
322b
invariance of
372
root locus method
127
from state–space models
323
465
323
T
Table lookup, in partial fraction expansion
22
tan command
144
Tangent method for PID controllers
156
25
Tank control systems
architecture
with piecewise constant inputs
Terminal penalties
tf command
tf2ss command
Time advance property
504
9b
404
30
44
265
266
79
330
16
535t
Time constant
analog control system design
132
z-domain function
174
Time delay, see Delay
Time function, in root locus method
135
This page has been reformatted by Knovel to provide easier navigation.
146
Index Terms
Links
Time response
convolution summation
32
convolution theorem
34
discrete-time systems
32
position control systems
133
steady-state regulator
423f
Time-delay
57
Time-limited functions
47
Torque vectors
133b
441b
Trace operations
eigenvalues
548
Leverrier algorithm
252
matrices
546
547b
Tracking errors
digital control systems
75
sampling period switching
516
servo problem
367
Tracking problem
423
Tracking time constant
509
Trajectories
equilibrium points
465
inertial control systems
414f
Transfer function matrices
264
BIBO stability
300
MATLAB
265
multivariable zeros
319
poles and zeros from
320
similar systems
282
system zeros
z-transfer
322b
277
Transfer functions
armature-controlled DC motor
67b
This page has been reformatted by Knovel to provide easier navigation.
Index Terms
Links
Transfer functions (Cont.)
BIBO stability
300b
cascades
60b
closed-loop
52b
combination systems
61
convolution theorem
35
DAC
498
duality
339
frequency response
furnace
301b
71
39
66b
MATLAB commands
79
PID controllers
153
pole assignment
389
programming
493
sampler effects
58
SIMULINK
80
from state–space representation
277
steady-state error
79b
transport delay lags
69
vehicle position control system
64
zero-order hold (ZOH)
57
Transfer of PID modes
153b
512
Transformations
bilinear, see Bilinear transformations
to controllable form
356
Fourier
45b
Laplace, see Laplace transforms
logarithmic
448
similarity
279
z-transforms, see Z-transforms
Transient response
analog control system design
132
This page has been reformatted by Knovel to provide easier navigation.
Index Terms
Links
Transient response (Cont.)
bilinear transformations
189
193
eigenvalues
359
363
low-pass filters
496
observer matrices
384
PD controllers
203
PI controllers
147
PID controllers
153
root locus design
136
system zeros
372
Transmission zeros
324
Transport lag, systems with
69
Transposition of matrices
538
Truncation, in ADCs
498
Turbojet engine example
Two degree-of-freedom control scheme
195
551
4
367
2-D.O.F. anthropomorphic manipulators
state–space equations
torque vectors
Type number, in digital control systems
241
441b
76b
U
Unbounded sequences, in final value theorem
28
Uncertainty equivalence principle
381
Unconstrained optimization
400
Uncontrollable modes
302
Uncontrollable systems
310b
Uniqueness proof, in Lyapunov equations
456
Unit impulse
convolution summation
z-transforms
Unit ramps
32
13b
36
77
This page has been reformatted by Knovel to provide easier navigation.
Index Terms
Links
Unit steps
14f
Unitary matrices
546
Unity feedback
75
57
77
78b
98
466f
468f
281
316
177
Unity matrices
543
Unobservable modes
313
Unobservable systems
318b
Unstable equilibrium points
459
Unstable systems
294
Upper triangular matrices
549
V
Van der Monde matrix
259
Vectors
column and row
gain
539b
360
MATLAB
30
matrix multiplication by
540
norms
552
perturbation
243
state
236
torque
441b
Vehicle position control systems
impulse response
64
ripple-free deadbeat controller
224b
z-domain digital control system design
177b
Velocity error constant
77
PD controllers
140
z-domain digital control system design
177
Velocity trajectory, in inertial control systems
414f
Verification of software
492
Very large scale integration (VLSI) technology
2
This page has been reformatted by Knovel to provide easier navigation.
108
Index Terms
Vortex equilibrium points
Links
465t
469f
W
Windup, integrator
509
W-plane, in frequency response design
206
208b
209b
Z
z-domain digital control system design
168
contours
171
direct
200
direct
213
finite settling time
218
frequency response
205
proportional control design
175
Z-domain pole locations
Z-domain root locus
179
93
165
zero command
323b
Zero matrices
543
Zero-input response
asymptotic stability
296b
controllability
302
discrete-time state equations
271
full-order observers
274
384b
linear state–space equations
248
pole placement
360
reduced-order observers
387f
361f
362f
Zero-order hold (ZOH)
frequency response
503
phase delay
503
PID controllers
199
transfer functions
57
This page has been reformatted by Knovel to provide easier navigation.
211b
Index Terms
Links
Zeros
bilinear transformations
189
190
decoupling
277
320
frequency response design
206
207
invariance of
372
multivariable systems
319
PI controllers
147
root locus method
127
single-axis milling machine
195
151
322b
state–space models
323
from transfer function matrix
320
444
Zero-state response
discrete-time state–space equations
271
linear state–space equations
248
z-transfer function from
277
Zero steady-state errors, see Steady-state errors
zgrid command
174
Ziegler-Nichols rules
158
ZOH blocks
489
199
200f
ZOH, see Zero-order-hold
z-domain digital control system design, analog
controllers, see Analog controller design
implementation
zpk command
ztrans command
101
279
30
z-transfer functions
armature-controlled DC motor
67b
cascades
60b
convolution theorem
35
DAC
498
furnace
66b
from state–space representation
277
This page has been reformatted by Knovel to provide easier navigation.
Index Terms
Links
z-transfer functions (Cont.)
transport delay systems
69
vehicle position control system
64
z-transforms
12
difference equations
31
final value theorem
28
frequency response
40
impulse response sequence
75
inversion
19
modified
37
PID controller incremental form
515
properties
15
sampled unit step input
77
servo problem
standard discrete-time signals
535t
368
13
state equations
272
tables of
533t
This page has been reformatted by Knovel to provide easier navigation.